Integrated energy system intelligent dispatching and control method and device

Abstract

The application provides a kind of integrated energy system intelligent scheduling and control method and device.The method obtains electric, gas, heat resource operation data, multi-scale hierarchical modeling is carried out according to time scale and space scale, establishes electric resource state space model, gas resource flow continuity model, heat resource heat balance model and multi-energy coupling characteristic constraint model;Deep auto-encoder network is used for feature extraction and dimension reduction representation;Through split hierarchical federated learning framework, collaborative training of multiple cognitive models is carried out, and global intelligent model is obtained;The neural architecture search network with hybrid bionic learning rule is constructed, the hierarchical scheduling target is set, and the hierarchical intelligent scheduling strategy is generated;Fault-tolerant control mechanism is constructed, error detection and correction of operation deviation are realized.The application realizes multi-time scale cooperation, collaborative learning under multi-device group privacy protection and fault-tolerant control with high reliability, significantly improves the operation efficiency and reliability of integrated energy system.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of energy management, and in particular to a comprehensive energy system intelligent scheduling and control method and device. BACKGROUND

[0002] Integrated Energy System (IES) as an important carrier to realize energy interconnection and improve energy utilization efficiency, has become the core component of modern energy system. Through the coordinated allocation and optimal scheduling of electricity, gas, heat and other multiple energy, the system can significantly improve the overall operation efficiency and economy.

[0003] At present, the scheduling control of comprehensive energy system mainly adopts centralized scheduling method based on mathematical optimization and distributed coordination method based on multi-agent. The centralized scheduling method establishes a mathematical model of the whole system, and uses linear programming, quadratic programming and other optimization algorithms to solve the optimal scheduling scheme, but it faces the problems of high computational complexity and poor real-time performance. The distributed coordination method divides the system into multiple subsystems, each subsystem makes independent decisions and realizes coordination through information interaction, which improves the scalability of the system, but has deficiencies in multi-scale collaboration and global optimality.

[0004] In recent years, with the rapid development of artificial intelligence technology, intelligent scheduling methods based on deep learning and reinforcement learning have gradually emerged. This kind of method learns the dynamic characteristics and scheduling rules of the system by constructing neural network model, which can handle the complex characteristics such as nonlinearity and uncertainty of the system. Typical implementation methods include load prediction based on deep neural network, real-time scheduling strategy generation based on reinforcement learning, etc. The core principle is to model the scheduling problem as a Markov decision process, and learn the optimal strategy through the interaction between the agent and the environment.

[0005] However, the existing technology still has significant deficiencies in dealing with the multi-scale characteristics of comprehensive energy system, multi-device collaborative cognition and system fault-tolerant control. Specifically, first, there is a lack of effective multi-time scale collaboration mechanism, which is difficult to balance the different needs of second-level frequency regulation, minute-level load adjustment and hour-level economic scheduling; second, there is a lack of privacy protection collaborative learning framework for each device group, which limits the improvement of the overall cognitive ability of the system; third, the existing scheduling control system lacks effective fault-tolerant mechanism and cannot cope with operation deviation and system failure. SUMMARY

[0006] The purpose of the present application is to overcome the above-mentioned deficiencies existing in the prior art, provide a comprehensive energy system intelligent scheduling and control method and device, which realizes the multi-time scale cooperation, collaborative learning under the privacy protection of multiple device groups and fault-tolerant control of high reliability of the comprehensive energy system through multi-scale hierarchical modeling, multi-group cognitive collaborative training under the split hierarchical federated learning framework, intelligent scheduling strategy generation of mixed bionic learning rule and fault-tolerant control based on confidence propagation ordered statistical decoding algorithm.

[0007] In order to achieve the above-mentioned purpose, the technical scheme adopted by the present application is as follows:

[0008] A comprehensive energy system intelligent scheduling and control method, comprising the following steps:

[0009] Obtain the electric, gas and heat resource operation data of the comprehensive energy system, and perform multi-scale hierarchical modeling on the operation data according to the second, minute and hour time scales and the local cluster and regional cluster spatial scales, to obtain an electric resource state space model, a gas resource flow continuity model, a heat resource heat balance model and a multi-energy coupling characteristic constraint model;

[0010] For the electric resource state space model, the gas resource flow continuity model, the heat resource heat balance model and the multi-energy coupling characteristic constraint model, a deep autoencoder network is used for feature extraction and dimension reduction representation to obtain resource representation data in a first dimension feature space;

[0011] The resource representation data in the first dimension feature space is subjected to multi-group cognitive model collaborative training through a split hierarchical federated learning framework, and parameter aggregation and global optimization of the multi-group cognitive model are performed based on the hierarchical architecture to obtain a global intelligent model with multi-group device collaborative cognitive ability;

[0012] A neural architecture search network with a mixed bionic learning rule is constructed in the global intelligent model, different scheduling targets of second-level frequency stability, minute-level load suppression and hour-level economic optimization are set, and optimal bionic learning rule combination search is performed in the neural architecture search network to generate and execute a hierarchical intelligent scheduling strategy;

[0013] The execution result of the hierarchical intelligent scheduling strategy is monitored in real time, a confidence propagation ordered statistical decoding BP-OSD algorithm is used to construct a fault-tolerant control mechanism, and error detection and correction are performed when the operation deviation of the comprehensive energy system exceeds a preset threshold.

[0014] Further, obtain the electric, gas, and heat resource operation data of the integrated energy system, perform multi-scale hierarchical modeling on the operation data according to the time scales of seconds, minutes, and hours and the spatial scales of local clusters and regional clusters, to obtain an electric resource state space model, a gas resource flow continuity model, a heat resource heat balance model, and a multi-energy coupling characteristic constraint model, including:

[0015] For the electric resource operation data, differential equations of state vectors and control inputs are used to establish a charging and discharging dynamic characteristic modeling of the electric resource, to obtain a relationship model between energy storage state vectors and control inputs;

[0016] For the gas resource operation data, continuity constraints of gas density and flow rate are used to model gas flow, to obtain a gas density and flow rate distribution model;

[0017] According to the heat resource operation data, the energy storage state vectors, the relationship model of control inputs, the gas density, and the flow rate distribution model, a heat transfer process modeling is performed using the balance relationship between mass heat capacity and temperature change, to construct a constraint relationship between electric power and gas power and heat power of multi-energy coupling equipment, to obtain the electric resource state space model, the gas resource flow continuity model, the heat resource heat balance model, and the multi-energy coupling characteristic constraint model.

[0018] Further, for the electric resource state space model, the gas resource flow continuity model, the heat resource heat balance model, and the multi-energy coupling characteristic constraint model, a deep auto-encoder network is used for feature extraction and dimension reduction representation, to obtain resource representation data in a first dimension feature space, including:

[0019] The electric resource state space model, the gas resource flow continuity model, the heat resource heat balance model, and the multi-energy coupling characteristic constraint model are preprocessed and normalized to obtain standardized multi-dimensional operation state data;

[0020] Based on the standardized multi-dimensional operation state data, an encoder network is constructed for nonlinear mapping of second dimension state data, to obtain a compressed intermediate representation vector, and the second dimension is higher than the first dimension;

[0021] A decoder network is used to reconstruct and verify the compressed intermediate representation vector, and the encoder parameters are optimized by minimizing the reconstruction error, to obtain the resource representation data in the first dimension feature space.

[0022] Further, the hierarchical architecture includes a first layer network deployed at a local client, a second layer network deployed at an edge server, and a third layer network deployed at a cloud server, and the split hierarchical federated learning framework is obtained by fusing split learning and hierarchical federated learning. For resource representation data of the first dimension feature space, the split hierarchical federated learning framework is used for collaborative training of multiple groups of cognitive models. On the basis of the hierarchical architecture, parameter aggregation and global optimization of the multiple groups of cognitive models are performed to obtain a global intelligent model with multiple groups of device collaborative cognitive capabilities, including:

[0023] For resource representation data of the first dimension feature space, preliminary feature learning is performed in the first layer network deployed at the local client, differential privacy and homomorphic encryption technology are used to protect data privacy, and an encrypted basic feature vector is obtained.

[0024] Based on the encrypted basic feature vector, deep representation learning is performed in the second layer network deployed at the edge server. An attention mechanism and a graph convolution network are used to learn cross-device associated features, and a regional collaborative representation result is obtained.

[0025] In the third layer network deployed at the cloud server, a federated averaging algorithm is used to aggregate and optimize the regional collaborative representation result, and a split learning cross-region coordination mode is used to obtain the global intelligent model with multiple groups of device collaborative cognitive capabilities.

[0026] Further, based on the encrypted basic feature vector, deep representation learning is performed in the second layer network deployed at the edge server. An attention mechanism and a graph convolution network are used to learn cross-device associated features, and a regional collaborative representation result is obtained, including:

[0027] In the second layer network deployed at the edge server, a four-to-six layer deep network is used to fuse features of the encrypted basic feature vector. A multi-head attention mechanism is used to calculate associated weights between different devices, and cross-device coupled features are obtained.

[0028] According to the cross-device coupled features, a graph convolution network is used to learn load transfer paths and fault propagation modes. A time series model is used to identify time-dependent relationships of the different devices, and a spatiotemporal association mode is obtained.

[0029] The spatiotemporal association mode is subjected to parameter aggregation and local optimization of the second layer network. A weighted average method is used to update parameters of the second layer network according to a client data amount weight and a loss function gradient, and the regional collaborative representation result is obtained.

[0030] Further, in the global intelligent model, a neural architecture search network with a hybrid biomimetic learning rule is constructed, different scheduling targets of second-level frequency stability, minute-level load suppression, and hour-level economic optimization are set, and optimal biomimetic learning rule combination search is performed in the neural architecture search network to generate and execute a hierarchical intelligent scheduling strategy, including:

[0031] In the global intelligent model, a reinforcement learning controller based on a long short-term memory network is used to construct the neural architecture search framework, and the current network architecture, historical performance indicators, and system operating conditions are used as state spaces to obtain a neural network architecture search state representation;

[0032] Based on the neural network architecture search state representation, the reinforcement learning controller is used to sample candidate architectures, and a candidate neural network architecture is generated through an action space of layer selection, activation function type, connection mode, and learning rule combination to obtain a diversified candidate neural network architecture set;

[0033] The diversified candidate neural network architecture set is evaluated for neural network architecture performance on a historical running data validation set, a reward signal is calculated through comprehensive indicators of scheduling accuracy, response speed, and energy efficiency, and a neural network architecture performance evaluation result is obtained;

[0034] According to the neural network architecture performance evaluation result, the policy gradient algorithm is used to optimize the controller policy parameters, the different scheduling targets of second-level frequency stability, minute-level load suppression, and hour-level economic optimization are used to guide the architecture search direction, and the optimal biomimetic learning rule combination is obtained;

[0035] For the optimal biomimetic learning rule combination, different frequency components are routed to the corresponding biomimetic learning layer through a time scale decomposer, and a hierarchical intelligent scheduling strategy is generated and executed through multi-layer information fusion using an attention mechanism.

[0036] Further, the hierarchical intelligent scheduling strategy includes a second-level frequency stability control strategy:

[0037] For the first frequency component in the hierarchical intelligent scheduling strategy, a frequency deviation signal is converted into a pulse firing rate proportional to the deviation amplitude through a Poisson encoder to obtain a frequency deviation pulse sequence;

[0038] Based on the frequency deviation pulse sequence, a leaky integrate-and-fire model of a spiking neural network is used to dynamically model the membrane potential, the membrane potential change is calculated through the time constant, resting potential, and input resistance of the membrane potential to obtain a pulse response signal;

[0039] The pulse response signal is threshold judged, a control pulse is generated when the membrane potential exceeds a firing threshold, and the membrane potential is reset to obtain a frequency adjustment instruction;

[0040] According to the frequency adjustment instruction, millisecond-level power regulation is performed on the generator set and energy storage device, and frequency deviation is eliminated through primary frequency control control;

[0041] After eliminating the frequency deviation, the frequency control state is monitored and fed back in real time, and the frequency stability error is controlled within the range of plus or minus 0.05 hertz.

[0042] Further, the layered intelligent scheduling strategy includes a minute-level load leveling control strategy:

[0043] The second frequency component in the layered intelligent scheduling strategy is extracted, the adaptive ability of load prediction is enhanced by using Hebb learning rule, the load change mode is obtained by simultaneously activating the reinforcement learning of neuron connection strength, and an adaptive load prediction model is obtained;

[0044] According to the adaptive load prediction model, a time window mechanism is introduced to learn the periodic pattern and mutation characteristics of the load, the prediction network parameters are trained by historical load data, and a load prediction result is obtained;

[0045] The load prediction result is selected by using a competitive learning mechanism, the key feature recognition ability is enhanced by lateral inhibition, the load prediction accuracy is improved to more than 95%, and a high-precision load prediction result is obtained;

[0046] Based on the high-precision load prediction result, the load deviation and fluctuation trend are calculated, the power compensation and peak clipping and valley filling are performed by using the energy storage system and controllable load, and a load leveling control strategy is obtained.

[0047] Further, the layered intelligent scheduling strategy includes a minute-level economic optimization control strategy:

[0048] The third frequency component in the layered intelligent scheduling strategy is identified, and a multi-objective optimization function is constructed by using an improved back propagation learning rule, including: minimizing the operation cost of electricity, gas and heat energy, minimizing carbon emissions and maximizing reliability;

[0049] The gradients of the multi-objective optimization function are weighted and combined by using a multi-objective gradient descent algorithm, and the weight coefficients are adaptively adjusted according to the operation state of the integrated energy system by using a dynamic weight adjustment strategy, to obtain optimal scheduling parameters;

[0050] The optimal scheduling parameters are verified under the constraint conditions, including device output limit, network transmission constraint, and energy balance constraint, and a feasible scheduling scheme is generated based on the verification result;

[0051] For the feasible scheduling scheme, an hour-level scheduling plan of different devices is formulated, including generator set start-stop timing, energy storage charging and discharging strategy, and cogeneration operation mode, to obtain an economic optimization scheduling strategy.

[0052] Further, the execution result of the hierarchical intelligent scheduling strategy is monitored in real time, and a belief propagation ordered statistical decoding BP-OSD algorithm is used to construct a fault-tolerant control mechanism for error detection and correction when the operation deviation of the integrated energy system exceeds a preset threshold, including:

[0053] The execution result of the hierarchical intelligent scheduling strategy is collected in real time, and different device states and energy flow data are encoded into check code words with error correction capability according to BCH code or LDPC code rules to obtain encoded data of the operation state of the integrated energy system;

[0054] For the encoded data, a belief propagation algorithm is used for error detection, and multiple rounds of iterative calculations are performed through factor graph representation and message passing from variable nodes to check nodes to obtain error position probability distribution;

[0055] Based on the error position probability distribution, an ordered statistical decoding algorithm is used to generate multiple candidate error correction schemes, the Euclidean distance and likelihood probability of each candidate error correction scheme and the received sequence are calculated, and the candidate error correction scheme with the maximum likelihood probability is selected for error correction to obtain the integrated energy system operation deviation correction instruction.

[0056] An integrated energy system intelligent scheduling and control device, comprising:

[0057] An acquisition module is configured to acquire electric, gas, and heat resource operation data of an integrated energy system, and to perform multi-scale hierarchical modeling on the operation data according to second-level, minute-level, and hour-level time scales and local cluster and regional cluster spatial scales to obtain an electric resource state space model, a gas resource flow continuity model, a heat resource heat balance model, and a multi-energy coupling characteristic constraint model.

[0058] A feature extraction and dimension reduction representation module is configured to perform feature extraction and dimension reduction representation on the electric resource state space model, the gas resource flow continuity model, the heat resource heat balance model, and the multi-energy coupling characteristic constraint model using a deep autoencoder network to obtain resource representation data in a first dimension feature space.

[0059] A collaborative training module is configured to perform multi-group cognitive model collaborative training on the resource representation data in the first dimension feature space through a split hierarchical federated learning framework, and to perform parameter aggregation and global optimization of the multi-group cognitive models based on the hierarchical architecture to obtain a global intelligent model with multi-group device collaborative cognitive ability.

[0060] The search module is configured to construct a neural architecture search network with a hybrid bionic learning rule in the global intelligent model, set different scheduling targets of second-level frequency stability, minute-level load suppression and hour-level economic optimization, and perform optimal bionic learning rule combination search in the neural architecture search network to generate and execute a hierarchical intelligent scheduling strategy.

[0061] The monitoring module is configured to monitor the execution result of the hierarchical intelligent scheduling strategy in real time, construct a fault-tolerant control mechanism by using a belief propagation ordered statistical decoding (BP-OSD) algorithm, and perform error detection and correction when the operation deviation of the integrated energy system exceeds a preset threshold.

[0062] The present application has the following advantages:

[0063] 1. By multi-scale hierarchical modeling and resource representation, the integrated energy system is modeled according to different time scales and spatial scales, and the characteristics of electricity, gas and heat are represented and learned, thereby improving the cognitive ability of the system to the coupling characteristics of multiple energy types.

[0064] 2. The split hierarchical federated learning framework is used for collaborative training of multiple cognitive models, realizing collaborative learning across devices and regions while protecting data privacy, and improving the overall cognitive ability and coordination optimization effect of the system.

[0065] 3. The neural architecture search network with the hybrid bionic learning rule is constructed, the optimal bionic learning rule combination is automatically searched, the hierarchical intelligent scheduling strategy suitable for different time scales is generated, and the comprehensive scheduling performance of the system under the multi-objective of second-level frequency stability, minute-level load suppression and hour-level economic optimization is improved.

[0066] 4. The fault-tolerant control mechanism is constructed by using the belief propagation ordered statistical decoding (BP-OSD) algorithm, the real-time detection and correction of the operation deviation of the integrated energy system are realized, and the reliability and robustness of the system are improved. BRIEF DESCRIPTION OF DRAWINGS

[0067] In order to more clearly illustrate the technical solutions of the embodiments of the present disclosure, the drawings needed in the embodiments will be briefly introduced below. The drawings herein are incorporated into the specification and form a part of the specification, which illustrate the embodiments consistent with the present disclosure, and are used to explain the technical solutions of the present disclosure together with the specification. It should be understood that the following drawings only show some embodiments of the present disclosure, and therefore should not be regarded as a limitation on the scope, and other related drawings can also be obtained by those skilled in the art without creative labor.

[0068] Figure 1A flow chart of a comprehensive energy system intelligent scheduling and control method provided for an embodiment of the present application is shown in FIG. 1.

[0069] Figure 2 An architecture diagram of a comprehensive energy system intelligent scheduling and control device provided for an embodiment of the present application is shown in FIG. 2. DETAILED DESCRIPTION

[0070] In order to make the objects, technical solutions and advantages of the embodiments of the present disclosure clearer, the following will clearly and completely describe the technical solutions in the embodiments of the present disclosure with reference to the drawings in the embodiments of the present disclosure. Obviously, the described embodiments are only part of the embodiments of the present disclosure and not all the embodiments. The components of the embodiments of the present disclosure described and shown in the drawings can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present disclosure provided in the drawings is not intended to limit the scope of the claimed present disclosure, but only represents selected embodiments of the present disclosure. Based on the embodiments of the present disclosure, all other embodiments obtained by those skilled in the art without creative work fall within the scope of protection of the present disclosure.

[0071] It should be noted that: similar reference numerals and letters represent similar items in the following drawings, so once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings.

[0072] The term "and / or" herein is only to describe an associated relationship, which means that there can be three relationships, for example, A and / or B can represent the following three cases: A exists alone, A and B exist together, and B exists alone. In addition, the term "at least one" herein means any one of a plurality or any combination of at least two of a plurality, for example, including at least one of A, B and C can mean including any one or more elements selected from the set consisting of A, B and C.

[0073] In order to make the objects, technical solutions and advantages of the embodiments of the present disclosure clearer, the following will clearly and completely describe the technical solutions in the embodiments of the present disclosure with reference to the drawings in the embodiments of the present disclosure. Obviously, the described embodiments are only part of the embodiments of the present disclosure and not all the embodiments. The components of the embodiments of the present disclosure described and shown in the drawings can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present disclosure provided in the drawings is not intended to limit the scope of the claimed present disclosure, but only represents selected embodiments of the present disclosure. Based on the embodiments of the present disclosure, all other embodiments obtained by those skilled in the art without creative work fall within the scope of protection of the present disclosure.

[0074] In order to make the objects, technical solutions and advantages of the embodiments of the present disclosure clearer, the following will clearly and completely describe the technical solutions in the embodiments of the present disclosure with reference to the drawings in the embodiments of the present disclosure. Obviously, the described embodiments are only part of the embodiments of the present disclosure and not all the embodiments. The components of the embodiments of the present disclosure described and shown in the drawings can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present disclosure provided in the drawings is not intended to limit the scope of the claimed present disclosure, but only represents selected embodiments of the present disclosure. Based on the embodiments of the present disclosure, all other embodiments obtained by those skilled in the art without creative work fall within the scope of protection of the present disclosure.

[0075] As shown in FIG. 1, the embodiments of the present application provide a comprehensive energy system intelligent scheduling and control method, which comprises the following steps: Figure 1

[0076] ​S1: Obtain the operation data of electric, gas and heat resources of the integrated energy system, and perform multi-scale hierarchical modeling on the operation data according to the time scales of seconds, minutes and hours and the spatial scales of local clusters and regional clusters to obtain an electric resource state space model, a gas resource flow continuity model, a heat resource heat balance model and a multi-energy coupling characteristic constraint model;

[0077] Step S1 is a multi-scale hierarchical modeling and resource representation stage. The integrated energy system (hereinafter referred to as the system) first obtains the operation data of electric, gas and heat resources in the integrated energy system. These data come from various sensors distributed in the entire energy network, including voltage, current, power, temperature, pressure and other parameters. The system performs hierarchical modeling on these operation data according to two dimensions of time and space. In the time dimension, it is divided into a second-level control layer, a minute-level regulation layer and a hour-level planning layer; in the spatial dimension, it is divided into a local device cluster and a regional coordination cluster. For electric resources, the system uses a state space model to describe the charge and discharge dynamic characteristics of electric resources such as energy storage devices, and establishes a relationship model between energy storage states and control inputs. For gas resources, the system establishes a gas flow model based on the continuity constraint of gas flow to describe the density and flow rate distribution of gas in the pipe network. For heat resources, the system uses a heat balance equation to describe the heat transfer process and establishes a heat balance model of heat resources. At the same time, the system also constructs the constraint relationship of multi-energy coupling devices, such as the constraint relationship between the electric power and the heat power of a combined heat and power unit. Through this multi-scale hierarchical modeling, the system can comprehensively capture the dynamic characteristics and operation rules of the integrated energy system at different time and space scales.

[0078] S2: For the electric resource state space model, the gas resource flow continuity model, the heat resource heat balance model and the multi-energy coupling characteristic constraint model, a deep autoencoder network is used for feature extraction and dimension reduction representation to obtain resource representation data in a first dimension feature space;

[0079] The system uses a deep auto-encoder network to extract features and reduce dimensionality representation for the electrical resource state space model, the gas resource flow continuity model, the thermal resource heat balance model, and the multi-energy coupling characteristic constraint model established in step S1. First, the system preprocesses the operation data of each type of resource model, including missing value filling, outlier processing, and data standardization, mapping data of different scales to a unified interval. Then, the system constructs an encoder network to non-linearly map the standardized multi-dimensional operation state data, compressing the high-dimensional operation state to a low-dimensional feature space to obtain the compressed intermediate representation vector. Next, the system constructs a symmetrical decoder network to remap the compressed intermediate representation vector to the original space, optimizing the network parameters by calculating the error between the reconstructed data and the original data. After training, the system extracts the intermediate representation vector output by the encoder as the low-dimensional feature representation of the resource, which contains the key information of the original operation data but greatly reduces the data dimension, facilitating subsequent cognitive model training and optimization calculation.

[0080] S3: For the resource representation data in the first dimension feature space, multiple cognitive models are collaboratively trained through a split hierarchical federated learning framework, and the parameters of the multiple cognitive models are aggregated and globally optimized based on the hierarchical architecture to obtain a global intelligent model with multi-device collaborative cognitive ability.

[0081] The resource representation data in the first dimension feature space refers to the expression of the characteristics, operating state, and mutual relationship of various energy devices in a multi-dimensional feature space through mathematical methods in a comprehensive energy system. The first dimension refers to the feature expression of the physical properties and operating characteristics of energy resources at the basic level. In this feature space, each energy device (such as generators, energy storage devices, heat pumps, etc.) is represented as a multi-dimensional vector, and each component of the vector describes the key characteristics of the device, such as capacity, efficiency curve, response time, operating cost, carbon emission intensity, etc. These representation data not only include static parameters, but also include dynamic characteristics of devices under different operating conditions, such as the change of the heat-to-power ratio of combined heat and power units at different load points, the efficiency change of energy storage devices at different charge and discharge depths, etc. In addition, the resource representation data also includes the topological structure information of the energy network, describing the physical connection relationship and energy transmission constraints between different devices. This multi-dimensional representation method enables the system to fully capture the complex characteristics and interactions of various energy devices, providing a mathematical basis for subsequent learning and optimization. Unlike traditional energy modeling, this representation method emphasizes the coupling characteristics and boundary conditions between devices, and can more accurately describe multi-energy complementarity and synergistic effects, serving as the data foundation for intelligent scheduling of multi-energy systems.

[0082] The split-hierarchical federated learning framework is an innovative distributed machine learning architecture designed to address the data silos and computational resource heterogeneity in multi-energy systems. While traditional federated learning mainly adopts horizontal or vertical partitioning strategies, the split-hierarchical federated learning framework combines both and adds a hierarchical structure. In this framework, "splitting" refers to dividing the learning task into multiple sub-tasks according to energy types and functional modules, with each sub-task handled by a specialized model. For example, power system state estimation, natural gas network pressure analysis, and heat network temperature prediction are all handled by specialized models. This splitting strategy allows each sub-model to focus on specific domain knowledge, improving learning efficiency. "Hierarchical" refers to the establishment of a multi-level learning structure from the device level to the system level. Bottom-level models perform local training within their respective devices or small areas, capturing micro-level energy characteristics and patterns. Middle-level models are responsible for coordination and optimization within similar energy systems, such as internal balancing and scheduling in power grids, and temperature and pressure control in heat networks. Top-level models are responsible for collaborative optimization between different energy systems, such as joint scheduling strategies for electricity-heat-gas systems. During the training process, each level of model collaborates through parameter sharing rather than raw data sharing, protecting sensitive data privacy. In terms of implementation, the system uses the federated averaging algorithm (FedAvg), introduces a weighting mechanism based on model performance and data quality, and designs an adaptive aggregation frequency strategy to ensure model convergence while reducing communication overhead. This framework is particularly suitable for complex systems with clear physical partitions and management boundaries, such as multi-energy systems. It can fully utilize the knowledge dispersed in various subsystems while protecting data privacy, and build a globally unified intelligent model.

[0083] The system adopts a split hierarchical federated learning framework to collaboratively train the resource representation data of the first dimension feature space obtained in step S2. The framework adopts a three-layer architecture design: local client layer, edge server layer and cloud server layer. In the local client layer, each local device cluster deploys a shallow neural network (first layer network) to perform preliminary feature learning, extracts the basic feature vector of the device running, and protects data privacy through differential privacy and homomorphic encryption technology. In the edge server layer, the system deploys an intermediate layer network (second layer network) for deep feature learning, receives feature vectors from multiple clients in the same region, and learns cross-device association features and time series patterns through attention mechanism and graph convolution network. The edge server aggregates and optimizes the spatio-temporal association pattern parameters, and updates the regional model parameters using weighted average method. In the cloud server layer, the system receives the representation results of each edge server, and performs global model aggregation and strategy generation through a deep network (third layer network). The global model is aggregated and optimized using federated averaging algorithm. Through this hierarchical split training method, the system protects the privacy of each device data while achieving collaborative improvement from local cognition to global intelligence, forming a global intelligent model with multi-device collaborative cognition capability.

[0084] In the embodiments of the present application, deep feature learning refers to the process of automatically extracting high-level abstract features from raw or preliminary processed data using deep neural networks. In the system, this technology is particularly suitable for processing high-dimensional, nonlinear, time-varying energy data. Specifically, the deep feature learning network deployed on the edge server adopts a multi-level neural network structure, usually including multiple convolution layers, pooling layers and fully connected layers. These network layers can extract abstract features from data step by step: low-level networks capture basic patterns such as short-term fluctuations in load curves, sudden changes in device state, etc.; middle-level networks identify more complex patterns such as daily load cycles, simple interactions between devices, etc.; high-level networks can understand system-level abstract concepts such as operating mode transitions, energy flow paths, etc. The unique advantage of deep feature learning is that it can adaptively discover patterns from data without the need for artificial pre-defined features, which enables the system to discover complex associations that traditional methods may overlook. In a multi-energy system, this learning method can capture the implicit relationships between different energy forms such as electricity, heat and gas, such as the electric-thermal coupling characteristics of combined heat and power units, the dynamic response characteristics of electric-gas conversion devices, etc. In addition, deep feature learning has strong generalization ability and can handle new operating scenarios and device combinations, which is particularly important for increasingly complex and changing energy systems. In actual deployment, the deep feature learning on the edge server usually adopts an incremental learning strategy, which can continuously update the model as new data arrives, maintaining accurate representation of system dynamic characteristics.

[0085] S4: constructing a neural architecture search network with hybrid biomimetic learning rules in the global intelligent model, setting up different scheduling targets of second-level frequency stability, minute-level load suppression, and hour-level economic optimization, and performing optimal biomimetic learning rule combination search in the neural architecture search network to generate and execute a hierarchical intelligent scheduling strategy;

[0086] The neural architecture search network with hybrid biomimetic learning rules is an innovative architecture that combines the learning mechanisms of biological neural systems and the automated design of artificial neural networks. The core idea of this technology is to draw on various learning rules observed in biological neural systems and find the optimal combination of neural network structure and learning rules for different scheduling tasks through automated architecture search methods. "Hybrid biomimetic learning rules" refer to the integration of multiple learning mechanisms inspired by biological neural systems, including but not limited to Hebbian learning (reflecting the principle of "co-activation of neurons enhances connection"), STDP (spike timing-dependent plasticity, simulating the influence of neuron activation timing on synaptic strength), homeostatic plasticity (stable plasticity, simulating the mechanism of self-regulating activity level of neurons), and neuromodulation (neurotransmitter modulation, simulating the role of chemical signals in the learning process). Unlike the single backpropagation rule in traditional deep learning, these biomimetic learning rules can better handle energy system data with strong temporal dependence and frequent multi-scale interactions. "Neural architecture search network" refers to the system's ability to automatically explore and optimize the structure of neural networks, including the number of layers, the number of neurons in each layer, the connection method, the activation function, and the learning rule. The system uses a reinforcement learning-based controller to generate candidate architectures and continuously optimizes the design through performance evaluation and feedback. To improve search efficiency, parameter sharing and progressive search strategies are introduced, significantly reducing the demand for computing resources. This neural architecture search network with hybrid biomimetic learning rules is particularly suitable for multi-energy system scheduling, which requires handling complex tasks involving multiple time scales and various physical processes. It can automatically design the most suitable neural network structure and learning mechanism for specific problem characteristics, improving the model's expression ability and generalization performance.

[0087] Step S4 is a hierarchical intelligent scheduling strategy generation stage. The system constructs a neural architecture search network with a hybrid bionic learning rule in the global intelligent model, and automatically discovers the optimal learning rule combination suitable for different scheduling levels. The system uses a reinforcement learning controller based on a long short-term memory network to construct a neural architecture search framework, and encodes the current network architecture, historical performance indicators, and system operating conditions as a state space. The controller samples candidate architectures based on state representation, and the action space includes network layer selection, activation function type, connection mode, and learning rule combination. The system evaluates the generated candidate architectures on the historical running data validation set, and calculates the reward signal through scheduling accuracy, response speed, energy efficiency, and other indicators. According to the evaluation results, the system optimizes the controller parameters using a policy gradient algorithm, and finally obtains the optimal bionic learning rule combination. The system routes the scheduling requirements of different frequency components to the corresponding bionic learning layer through a time scale decomposer: high-frequency components are routed to the spiking neural network layer for frequency stability control, medium-frequency components are processed by the Hebb learning layer for load suppression control, and low-frequency components are handed over to the back propagation layer for economic optimization control. Each layer fuses information through an attention mechanism to form a complete hierarchical intelligent scheduling strategy, realizing multi-objective collaborative scheduling of second-level frequency stability, minute-level load suppression, and hour-level economic optimization.

[0088] Among them, the optimal bionic learning rule combination search refers to the process of automatically finding and determining the learning rule combination most suitable for a specific scheduling task in a neural architecture search network with mixed bionic learning rules. This search process aims to find learning rule configurations that can maximize performance for scheduling targets at different time scales (second-level frequency stability, minute-level load smoothing, and hour-level economic optimization). The search process adopts a hierarchical strategy: first, the system predefines a series of candidate learning rules, each of which simulates a specific learning mechanism in the biological nervous system, such as forward synaptic plasticity, feedback inhibition regulation, reward-modulated learning, etc. Then, the system assigns different learning rules to different components of the neural network (such as different layers and different types of connections), forming a learning rule combination. Next, through a policy gradient-based reinforcement learning method, the system evaluates the performance of different combinations on specific scheduling tasks and gradually optimizes the combination strategy. During the evaluation process, the system considers multiple performance indicators, including prediction accuracy, computational efficiency, convergence speed, and robustness to noise, etc. Research shows that for tasks such as second-level frequency stability that require fast response, the combination of STDP and local inhibition rules performs best, as this combination can quickly capture short-term dependencies in time series; for minute-level load smoothing tasks, the combination of Hebbian learning and homeostatic plasticity is effective, as it can balance system response and stability; and for hour-level economic optimization tasks, the reinforcement learning rule combined with neuromodulation and reward signals performs best, as it can effectively handle long-term cumulative revenue optimization problems. Through this optimal bionic learning rule combination search, the system can customize the most suitable learning mechanism for scheduling tasks at different time scales, significantly improving the flexibility and adaptability of the multi-energy system scheduling.

[0089] S5: Real-time monitoring of the execution results of the hierarchical intelligent scheduling strategy, constructing a fault-tolerant control mechanism using the belief propagation ordered statistical decoding BP-OSD algorithm, and performing error detection and correction when the operation deviation of the integrated energy system exceeds the preset threshold.

[0090] The BP-OSD algorithm is an innovative error detection and correction technique that builds a fault-tolerant control mechanism to ensure the reliability of the scheduling strategy execution process in multi-energy systems. This technology migrates advanced error correction coding theories originally used in communication systems to the field of energy system control, building a complete fault-tolerant control framework. The BP-OSD algorithm combines two powerful decoding techniques: Belief Propagation (BP) and Ordered Statistics Decoding (OSD). Belief Propagation is an algorithm for probabilistic inference on factor graphs that can effectively handle complex probability models with local structure. In multi-energy systems, the BP algorithm is used to monitor the correlation between system state variables (such as voltage, pressure, temperature, etc.). When the observed state deviates from the expected pattern, the algorithm can calculate the most likely source of error. However, the BP algorithm may not converge to the correct solution in some cases (such as error accumulation or strong correlation errors). To solve this problem, the system introduces OSD as a supplementary mechanism. The OSD algorithm first sorts the state variables according to their confidence levels, and then systematically tests the least reliable variable combinations to find the most likely error pattern. This hybrid strategy of "soft information hard decision" greatly improves the success rate of error detection and correction. In practical applications, the system encodes the physical constraints and device characteristics of the multi-energy network into a check matrix, and any state that violates these constraints is considered a potential error. When the detected deviation exceeds the preset threshold, the BP-OSD algorithm will immediately start, analyze the deviation pattern and infer the most likely cause of the error (such as sensor failure, actuator malfunction or external interference, etc.). Based on the inference results, the system can automatically implement targeted correction measures, such as adjusting control parameters, switching to backup devices, or re-planning the scheduling strategy, etc. This fault-tolerant control mechanism significantly improves the robustness and self-recovery ability of multi-energy systems in the face of uncertainty and faults, ensuring the reliable execution of the scheduling strategy.

[0091] Step S5 is a fault-tolerant control and real-time feedback stage. The system adopts the belief propagation ordered statistical decoding BP-OSD algorithm to construct a multi-level fault-tolerant control mechanism, and monitors the operation state of the integrated energy system in real time and detects and corrects errors. The system collects key operating parameters in real time at a high frequency, removes measurement noise through a filter, and obtains high-precision state estimation values. The system converts these data into code words with error correction capability according to the coding rules, forming the coded data of the system operation state. When the operation deviation is detected to exceed the preset threshold, the system starts the fault-tolerant mechanism. First, the system uses the belief propagation algorithm to detect errors in the coded data, and through factor graph representation and message passing, it performs multiple rounds of iterative calculations to obtain the posterior probability of each bit position and determine the most likely error location. Then, based on the error position probability distribution, the system uses the ordered statistical decoding algorithm to generate multiple candidate error correction schemes, calculates the distance and likelihood probability of each scheme from the received sequence, and selects the optimal scheme for error correction. Finally, the system generates specific control instructions based on the error correction results, and issues them to each execution device through the communication network to correct the system operation deviation. Through this fault-tolerant control mechanism, the system can quickly locate and correct errors after detecting abnormalities, control the system operation deviation within a reasonable range, and significantly improve the reliability and robustness of the integrated energy system.

[0092] In this embodiment, the operation data of the electric, gas, and heat resources of the integrated energy system are obtained, and the operation data are modeled in multiple scales and layers according to the time scales of seconds, minutes, and hours and the spatial scales of local clusters and regional clusters, to obtain an electric resource state space model, a gas resource flow continuity model, a heat resource heat balance model, and a multi-energy coupling characteristic constraint model, including:

[0093] S1.1: For the electric resource operation data, the differential equation of the state vector and the control input is used to establish the electric resource charging and discharging dynamic characteristic modeling, to obtain a relationship model between the energy storage state vector and the control input;

[0094] In this step, the system first collects the operation data of the electrical resources from various sensors and measurement devices, including the state of charge, charging and discharging power, voltage, current, and other key parameters of the energy storage devices. The system uses differential equations of state vectors and control inputs to describe the dynamic characteristics of electrical resources, especially the charging and discharging behavior of energy storage devices. In this model, the state vector includes the state of charge (SOC), power output, device temperature, and other state variables of the energy storage device, which can fully reflect the operating state of the device; the control input represents the charging and discharging instructions issued to the device, such as charging power, discharging power, etc. The system establishes a relationship model between the state vector of the energy storage and the control input by analyzing the change rule of the state vector over time. This model can accurately describe the state change process of the energy storage device after receiving different control instructions, providing an important basis for subsequent optimization scheduling. For example, for a battery energy storage system, the model can describe the change of battery capacity, internal temperature, and remaining life under different charging and discharging powers, so as to meet the system requirements and protect the energy storage device and prolong its service life during scheduling.

[0095] Specifically, the system performs state space modeling on electrical resources, using differential equations dx / dt = Ax + Bu to describe the charging and discharging dynamic characteristics of electrical resources, where x is the state vector of the energy storage, representing the state of charge (SOC), power output, and other parameters of the energy storage device; u is the control input, representing the charging and discharging instructions; A is the system matrix, describing the internal state transition relationship of the system; B is the control matrix, describing the influence of control input on the system state.

[0096] Among them, the differential equation of state vector and control input is a mathematical tool to describe the dynamic behavior of integrated energy system, which expresses the relationship between the physical state of the system and the control decision in an accurate mathematical form. In integrated energy systems, the state vector refers to a set of variables that can fully describe the state of the system at a certain time. Typical state variables include frequency, voltage, and phase angle in power systems, temperature, pressure, and flow in thermal systems, and pressure, flow, and composition in pneumatic systems. The number of these state variables can be very large, and in large integrated energy systems it can reach thousands or even tens of thousands. Control input refers to variables that can be directly adjusted by system operators or automatic control systems, such as generator output set value, transformer tap position, valve opening, pump speed, etc. The differential equation describes the law of change of these state variables over time and how control input affects this change. Specifically, it shows the functional relationship between the rate of change of state variables (i.e. derivative) and the current state and control input. For example, the rate of change of the frequency of the power system is related to power imbalance, system inertia and damping coefficient; the rate of change of the temperature of the thermal system is related to heat input, heat loss and heat capacity. These differential equations are usually nonlinear, reflecting the complex physical processes and interactions in energy systems. By solving these differential equations, we can predict the future behavior of the system under a given control strategy, evaluate the effectiveness of the control strategy, or conversely, find the optimal control strategy that can make the system achieve the desired state. In multi-energy system scheduling, these differential equations are the theoretical basis for building simulation models and predictive controllers, supporting the whole process from physical understanding to intelligent decision-making.

[0097] S1.2: For gas resource operation data, gas density and flow continuity constraints are used to model gas flow, and gas density and flow distribution models are obtained;

[0098] The system collects operation data of gas resources, including gas pressure, flow rate, temperature and other parameters in the natural gas pipeline network. Based on the principles of fluid mechanics, the system uses the continuity constraint of gas density and flow rate to establish a gas flow model. This model takes into account the basic physical laws of gas flow in the pipeline network, especially the principle of mass conservation, i.e. the mass of gas flowing into the pipeline should be equal to the mass of gas flowing out plus the change of gas mass stored in the pipeline. The system analyzes the distribution law of gas density and flow rate in the pipeline network and establishes a gas density and flow rate distribution model. This model can describe the flow behavior of gas under different pipeline parameters (such as diameter, length, roughness) and external conditions (such as ambient temperature, pressure), providing an accurate physical model basis for gas network scheduling. Through this model, the system can predict the flow direction, flow rate and pressure distribution of gas under different scheduling schemes, ensuring that the gas network operates within a safe range while meeting the gas demand of each user. In addition, the model also considers the pressure drop effect and temperature change during long-distance gas transportation, further improving the accuracy and applicability of the model.

[0099] Specifically, the system models the gas resources using the flow continuity equation where ρ is the gas density and v is the flow rate vector. This equation describes the flow characteristics of gas in the pipeline network. By solving this equation, the gas density and flow rate distribution at different nodes can be obtained, providing a basic model for gas network scheduling.

[0100] S1.3: According to the thermal resource operation data, the energy storage state vector, the control input relationship model, the gas density and the flow rate distribution model, the heat transfer process is modeled using the balance relationship of mass heat capacity and temperature change, and the constraint relationship between electric power, gas power and heat power of multi-energy coupling equipment is constructed, to obtain the electric resource state space model, the gas resource flow continuity model, the thermal resource heat balance model and the multi-energy coupling characteristic constraint model.

[0101] Based on the electrical resource state vector and control input relationship model obtained in the first two steps, the gas density and flow rate distribution model, and the operation data of thermal resources, the system uses the balance relationship between mass heat capacity and temperature change to model the heat transfer process. The heat balance model describes the dynamic change law of thermal energy in the storage, transmission and conversion process, and considers key factors such as heat capacity, heat loss and heat exchange efficiency. For a district heating system, the model describes the heat transfer process between the heat source and the user, including the relationship between parameters such as hot water temperature, flow rate and supply and return water temperature difference. On this basis, the system further builds the constraint relationship between electrical power and gas power and thermal power of multi-energy coupling equipment, forming a complete multi-energy coupling characteristic constraint model. This model is particularly suitable for multi-energy coupling equipment such as combined heat and power units, gas turbines, and electric-gas-heat tri-generation systems, and describes the conversion law and constraint conditions between different energy forms. For example, for a combined heat and power unit, the model establishes a constraint relationship of the electrical-thermal ratio, describing the coupling characteristics of electrical and thermal power output under different load conditions; for a gas turbine, the model establishes the relationship between gas consumption and electrical power output, as well as the efficiency constraint of waste heat recovery. Through these models, the system can fully grasp the mutual influence and conversion relationship between electric-gas-heat multi-energy, providing a solid theoretical basis for multi-energy collaborative optimization.

[0102] Specifically, the system uses the heat balance equation mcΔT = Q to describe the heat transfer process, where m is the mass, c is the specific heat capacity, ΔT is the temperature change, and Q is the heat. For multi-energy coupling equipment, the constraint relationship between electrical power Pe, gas power Pg and thermal power Qh is established as Pe = f(Pg, Qh), for example, for a combined heat and power unit, the constraint relationship of the electrical-thermal ratio η = Pe / Qh can be established.

[0103] In this embodiment, for the electrical resource state space model, the gas resource flow continuity model, the thermal resource heat balance model and the multi-energy coupling characteristic constraint model, a deep autoencoder network is used for feature extraction and dimension reduction representation to obtain resource representation data in a first dimension feature space, including:

[0104] S2.1: Data preprocessing and normalization are performed on the electrical resource state space model, the gas resource flow continuity model, the thermal resource heat balance model and the multi-energy coupling characteristic constraint model to obtain standardized multi-dimensional operation state data;

[0105] The system first conducts comprehensive preprocessing on the data of the electrical resource state space model, the gas resource flow continuity model, the thermal resource heat balance model, and the multi-energy coupling characteristic constraint model obtained from S1. This preprocessing process includes three main tasks: missing value filling, outlier processing, and data standardization. For missing values, the system uses different filling strategies according to the data type, such as forward filling or linear interpolation for time series data, and nearest neighbor filling or mean filling for state data. For outliers, the system uses a sliding window method to calculate the mean and standard deviation of the data, identifies data points outside the normal range, and processes them through limiting or replacing. The most critical is the data standardization process, which uses the Z-score standardization method to convert the data to a standard normal distribution with a mean of 0 and a standard deviation of 1, or the Min-Max normalization method to map the data to the 0-1 interval. Through this standardization process, the system solves the huge difference in dimension and numerical range of different energy data, such as power data usually in megawatts, while temperature data is in degrees Celsius. The standardized multi-dimensional running state data has consistent numerical characteristics, which provides high-quality input data for subsequent deep learning models, avoids the problem of certain features dominating model training due to large numerical values, and ensures equal learning of all types of features by the model.

[0106] Specifically, the system preprocesses the running data of various resource models, including missing value filling, outlier processing, and data standardization. Standardization uses the Z-score standardization method x' = (x - μ) / σ or the Min-Max normalization method x' = (x - xmin) / (xmax - xmin) to map data of different scales to a unified interval, facilitating subsequent feature extraction.

[0107] S2.2: Based on the standardized multi-dimensional running state data, an encoder network is constructed for nonlinear mapping of second-dimensional state data, obtaining a compressed intermediate representation vector, the second dimension being higher than the first dimension;

[0108] S2.2 is the stage of constructing the encoder network for feature space mapping. Based on the standardized multi-dimensional operating state data obtained in S2.1, the system constructs a multi-layer encoder neural network structure. This encoder adopts a fully connected neural network design, starting from the input layer, gradually compressing the data dimension through multiple hidden layers. The specific structure usually includes an input layer with node number equal to the dimension of the original data (possibly hundreds or thousands of features); followed by multiple hidden layers with decreasing node numbers, for example, possibly starting from 512 nodes, and then decreasing to 256, 128, 64 nodes; the node number of the final output layer corresponds to the target low-dimensional feature space, usually reduced to 5%-10% of the original dimension. Nonlinear activation functions such as ReLU or Sigmoid function are used between each layer to enhance the network's ability to capture complex nonlinear relationships. This layer-by-layer compression design enables the encoder to learn the key features and internal patterns in the data, mapping high-dimensional original operating state data to low-dimensional feature space, forming a compressed intermediate representation vector. This process is essentially a nonlinear dimensionality reduction of the original data, preserving the most important information in the data and removing redundancy and noise. It is worth noting that the feature extraction process here compresses the original data from the second dimension (usually higher dimension) to the first dimension (lower dimension), making the data more compact and informative.

[0109] Specifically, the system constructs an encoder network to perform nonlinear mapping on the standardized multi-dimensional operating state data. The encoder consists of multiple layers of fully connected neural networks, with n input nodes in the first layer (corresponding to the original data dimension), followed by multiple hidden layers (such as 512-256-128-64 nodes) to gradually compress the dimension, and finally output a d-dimensional intermediate representation vector (d is much smaller than n), achieving a compressed representation of the data.

[0110] It is necessary to note that S2.2 describes a process of data dimension reduction and feature extraction, which is a key step in the intelligent scheduling of multi-energy systems. Standardized multi-dimensional operating state data refers to system operating data that has been pre-processed (such as normalization, denoising, and outlier processing). These data come from sensors and measurement devices distributed throughout the multi-energy system and record the system's operating state at different time points. These raw data have extremely high dimensions and may contain redundant information and noise, which can lead to excessive computational burden and decreased model performance if used directly. The encoder network is a neural network structure specifically designed to map high-dimensional data to low-dimensional space while preserving key information in the data. In this process, the second-dimensional state data refers to the feature set obtained after preliminary processing of the original operating data, which already contains a certain degree of abstraction and organization, but the dimension is still high. Non-linear mapping emphasizes that this process is not a simple linear transformation (such as principal component analysis), but rather utilizes the non-linear activation function and multi-layer structure of neural networks to capture complex non-linear relationships in the data. This non-linear mapping can identify hidden patterns and features that linear methods cannot discover. Through the processing of the encoder, the system ultimately obtains a compressed intermediate representation vector, which is a low-dimensional vector but condenses the core information in the original data. The second dimension is higher than the first dimension, indicating that during the processing, the system first maps the data to a higher-dimensional space (by increasing feature combinations or introducing auxiliary variables), and then compresses it. This strategy of first increasing the dimension and then reducing it can make it easier for the system to discover non-linear patterns in the data, similar to the kernel trick in kernel methods. This compressed intermediate representation vector not only greatly reduces the computational load of subsequent processing, but also provides a more structured and informative representation, laying the foundation for the state understanding and decision optimization of multi-energy systems.

[0111] S2.3: Adopting a decoder network to reconstruct and verify the compressed intermediate representation vector, optimizing the encoder parameters by minimizing the reconstruction error, and obtaining resource representation data in the first-dimensional feature space.

[0112] S2.3 is the stage of reconstruction verification and optimization through the decoder network. In this step, the system constructs a decoder network symmetric to the encoder structure, which is used to map the compressed intermediate representation vector obtained in S2.2 back to the original data space. The structure of the decoder network is a mirror inversion of the encoder, starting from the low-dimensional intermediate representation, gradually expanding the dimension through multiple hidden layers, and finally recovering to the same dimension as the original data. For example, if the structure of the encoder is to compress from the original dimension through 512-256-128-64 node layers, then the structure of the decoder is to expand from 64 nodes through 128-256-512 node layers, and finally recover to the original dimension. The system inputs the intermediate representation vector output by the encoder into the decoder to obtain the reconstructed data, and then calculates the reconstruction error between the reconstructed data and the original input data, usually using mean square error as the loss function. The system minimizes this reconstruction error through the backpropagation algorithm, while optimizing the parameters of the encoder and the decoder. This training process is unsupervised and does not require additional labeled data, relying only on the structural information of the data itself for learning. When the reconstruction error is reduced to a satisfactory level, the system extracts the intermediate representation vector output by the encoder as the final resource representation data of the first-dimensional feature space. These representation data, although greatly reduced in dimension, retain the key information in the original data and can effectively represent the characteristics and status of different energy resources, providing high-quality feature input for subsequent cognitive model training and intelligent scheduling strategy generation.

[0113] Specifically, the system constructs a symmetric decoder network (such as 64-128-256-512 nodes) to remap the compressed d-dimensional intermediate representation vector to the original n-dimensional space, and optimizes the network parameters by calculating the mean square error loss L = ||x - x'|| 2 between the reconstructed data and the original data. After training, the intermediate representation vector output by the encoder is extracted as the low-dimensional feature representation of the resource, which is used for subsequent cognitive model training.

[0114] The resource representation data of the first dimension feature space refers to a data set formed after the basic characteristics of various energy resources and equipment are mathematically described in the system. The "first dimension" here refers to the most basic and direct feature level of the energy system, which focuses on the inherent attributes and basic operating characteristics of a single resource or equipment, rather than higher-level system behavior or abstract concepts. In this feature space, each energy resource or equipment is represented as a multi-dimensional data point, with the dimensions determined by various attributes describing the resource. For example, for a generator set, these attributes may include maximum output, minimum output, ramp rate, start-up time, thermal efficiency curve, fuel type, emission coefficient, etc.; for energy storage equipment, they may include capacity, charge and discharge efficiency, maximum charge and discharge power, self-discharge rate, cycle life, etc.; for thermal equipment, they may include thermal power range, energy conversion efficiency, temperature output range, etc. These representation data include both static parameters (such as equipment rated values, physical limitations, etc.) and dynamic characteristics (such as the relationship between efficiency and load, performance degradation over time, etc.). The collection of resource representation data is usually based on equipment specification parameters, historical operation data analysis, professional test results and physical model derivation, etc. In the intelligent scheduling of multi-energy systems, these resource representation data of the first dimension feature space constitute the basic knowledge base, providing raw information for higher-level system modeling and optimization decisions. It is worth noting that this representation method unifies different types of energy resources (such as power equipment, thermal equipment, pneumatic equipment, etc.) into a standardized feature space, facilitating subsequent collaborative analysis and optimization, which is the data foundation for the collaborative scheduling of multi-energy systems. Through analysis and learning of these resource representation data, the system can understand the characteristics and limitations of different energy resources, and thus develop technically feasible and economically efficient scheduling strategies.

[0115] In the present embodiment, the hierarchical architecture includes a first layer network deployed at local clients, a second layer network deployed at edge servers and a third layer network deployed at cloud servers. The split hierarchical federated learning framework is obtained by fusing split learning and hierarchical federated learning. For the resource representation data of the first dimension feature space, the split hierarchical federated learning framework is used for collaborative training of multiple cognitive models. The parameters of the multiple cognitive models are aggregated and globally optimized based on the hierarchical architecture to obtain a global intelligent model with multiple device collaborative cognitive capabilities, including:

[0116] S3.1: For the resource representation data of the first dimension feature space, preliminary feature learning is performed in the first layer network deployed at the local clients. Differential privacy and homomorphic encryption technology are used to protect data privacy, and encrypted basic feature vectors are obtained.

[0117] In this step, the system deploys a shallow neural network on each local device cluster for preliminary feature learning. Each local client is usually a specific energy subsystem, such as a regional power system, gas system, or heat system. These local clients independently collect and process the operation data of their respective devices, including power voltage, power, gas flow, pressure, heat temperature, flow, and other parameters. The shallow neural network deployed by the system on each client is usually composed of 2-3 layers of fully connected networks that receive the resource representation data processed by S2 as input and extract the basic feature vectors of device operation. Since each client only processes local data, the computational burden is lighter, and the system can run efficiently even on edge devices with limited computing resources. To protect data privacy, the system applies two layers of security mechanisms in the local feature learning process: first, differential privacy technology, which adds carefully designed noise to the training data or model parameters to ensure that even if it is deduced, the original sensitive data cannot be recovered; second, homomorphic encryption technology, which allows direct computation and processing of data in an encrypted state, and data remains encrypted throughout transmission and processing. This privacy protection design allows each energy unit to safely participate in collaborative learning without worrying about the risk of commercial sensitive data leakage, greatly enhancing the feasibility of the system in practical applications.

[0118] Specifically, the system deploys a 2-3 layer fully connected network on the local client for preliminary feature learning. The local client collects raw operation data such as voltage, current, temperature, and pressure from sensors, extracts the basic feature vectors of device operation through the first layer network, including power features, efficiency features, and state features, and the feature dimension is usually compressed to 10-20% of the original data. To protect data privacy, the system uses differential privacy technology to add calibration noise to the feature vectors and uses homomorphic encryption technology to encrypt the feature vectors, ensuring the security of data during transmission and processing.

[0119] S3.2: Based on the encrypted basic feature vectors, a second layer network deployed on the edge server is used for deep representation learning, using attention mechanisms and graph convolution networks to learn cross-device correlation features, resulting in regional collaborative representation results.

[0120] Attention mechanism and graph convolutional network learn cross-device correlation features, which is an innovative combination of deep representation learning techniques, specifically designed to capture the complex interaction between devices in a multi-energy system. Attention mechanism is a neural network component inspired by human cognition, which allows the model to "focus" on the most relevant parts of the input data. In this system, the importance of different devices at different times varies, for example, during periods of large fluctuations in renewable energy output, the state of energy storage devices may be more critical; while in high load periods, the adjustment capacity of traditional generating units may be more important. Attention mechanism calculates the importance weight of input features, allowing the model to dynamically adjust the attention to different devices and data at different time points, thereby improving the sensitivity to key information. In specific implementation, the system uses multi-head self-attention mechanism, which can evaluate the correlation strength between devices from multiple perspectives (such as time correlation, functional similarity, physical connection relationship, etc.). Graph convolutional network (GCN) is a deep learning method specially designed for processing graph structure data. Multi-energy systems naturally have graph structure characteristics, where nodes represent various energy devices, and edges represent the physical connection or energy flow relationship between devices. GCN can directly perform convolution operations on this graph structure, effectively capturing the spatial relationship and interaction mode between devices. Unlike traditional convolutional networks, the convolution kernel of GCN considers the neighborhood structure of nodes, allowing information to propagate along the actual topology of the energy network. Through multi-layer graph convolution, the system can gradually capture the correlation between devices at a longer distance, for example, how changes in regional power load affect the operation of remote devices in the heat network through intermediate devices such as electric-thermal conversion devices. By combining attention mechanism and graph convolutional network, the system forms an "attention-guided graph convolution" architecture, which can adaptively determine the important paths of information propagation and accurately capture the complex interaction patterns across devices and energy forms, ultimately forming a regional collaborative representation result, providing in-depth system understanding for the collaborative scheduling of multi-energy systems.

[0121] According to the cross-device coupling feature, the graph convolutional network is used to learn the load transfer path and the fault propagation mode, which is a breakthrough technology application in the intelligent dispatching of integrated energy systems. The graph convolutional network (GCN) is a deep learning architecture designed specifically for processing graph-structured data, which can perform convolution operations directly on graph data in non-Euclidean space. In a multi-energy system, various energy devices and their interconnections naturally form a complex graph structure, where nodes represent power generators, substations, cogeneration plants, energy storage devices, etc., and edges represent physical connections such as power lines, heat pipe networks, and natural gas pipelines. The core idea of the graph convolutional network is that the features of each node are not only affected by its own attributes but also by its neighboring nodes, which is highly consistent with the interactions between devices in a multi-energy system. When learning the load transfer path, GCN can capture how load changes are transferred from one device to adjacent devices and further affect the entire system. For example, when the power load suddenly increases, the grid frequency decreases, which may trigger the cogeneration unit to increase power output, thereby changing the heat supply of the heat network, and ultimately affecting the heating conditions of end users. This complex load transfer chain is gradually unfolded and learned by GCN through multiple layers of graph convolution operations. For fault propagation patterns, GCN also performs well. System failures often exhibit a chain reaction characteristic, where a failure in one device can trigger a cascade effect, affecting multiple related devices. By training on historical fault data, GCN can identify typical propagation paths and impact ranges of faults, providing decision support for fault prevention and emergency response. Another advantage of GCN is its inductive ability, which allows it to make reasonable predictions even for network topologies or device combinations that have not appeared in the training data, which is particularly important for evolving energy systems. In addition, GCN supports heterogeneous graph processing, allowing it to consider different types of nodes (such as power devices and heat devices) and edges (such as physical connections and control relationships), fully capturing the complex structural characteristics of multi-energy systems.

[0122] S3.2 is the deep feature learning stage of the edge server layer. Edge servers are usually deployed in regional energy management centers and are responsible for coordinating the learning process of multiple local clients within the same region. This stage can be further divided into three sub-steps: spatiotemporal correlation pattern learning (S3.2.1), parameter aggregation and local optimization (S3.2.2), and regional model updating (S3.2.3). This includes:

[0123] S3.2.1: In the second layer network deployed in the edge server, a four to six layer deep network is used to fuse the encrypted basic feature vectors, and the correlation weights between different devices are calculated through a multi-head attention mechanism to obtain cross-device coupling features;

[0124] Specifically, the edge server deploys a 4-6 layer deep network to receive the basic feature vectors uploaded by multiple clients in the same region, and performs feature fusion on these features. The system uses a multi-head attention mechanism to calculate the correlation weight between different devices, and the attention calculation formula is a ij = softmax(Q i K j T / d), where Q i and K j are the query vector and key vector of device i and device j, respectively. Through the multi-head attention mechanism, the system can automatically discover the correlation pattern between devices and form cross-device coupling features.

[0125] S3.2.2: According to the cross-device coupling features, learn the load transfer path and fault propagation mode using a graph convolution network, identify the time-dependent relationship of the different devices running through time series modeling, and obtain the spatio-temporal correlation pattern;

[0126] S3.2.1 is the spatio-temporal correlation pattern learning process. In this step, the edge server receives feature vectors uploaded by multiple clients in the same region, and performs deep representation learning through a more complex intermediate layer network structure. This network mainly consists of two parts: attention mechanism and graph convolution network. The attention mechanism can automatically identify and highlight important correlations between different devices, for example, when the operating state of a certain power device changes, it may have a significant impact on the connected heat device, the attention mechanism will give this correlation a higher weight. The graph convolution network abstracts the energy devices in the region as nodes and the energy flow relationship between the devices as edges to form an energy network topology graph structure. Through learning of this graph structure, the system can capture the spatial correlation pattern between devices, such as how the electric-thermal conversion device affects the coordinated operation of the power grid and the heat grid. At the same time, the system also uses a time series learning module (such as a long short-term memory network) to capture the time pattern of energy data, including periodic changes, trend changes, and sudden event patterns. Through this spatio-temporal combined learning method, the edge server can form a more comprehensive and in-depth understanding of the regional energy system, and discover complex correlation rules across devices and across energy types.

[0127] In one embodiment, the time-dependent relationship of different devices is identified by time series modeling, and the spatio-temporal correlation pattern is obtained, which is an important supplement to the spatial analysis capability of the graph convolution network. The devices in the multi-energy system not only have spatial interactions, but also have complex time-dependent relationships. Time series modeling aims to capture these dynamic characteristics that evolve over time. Specifically, the system adopts a fusion scheme of multiple time series analysis techniques. Long short-term memory network (LSTM) is used to capture long-term dependencies, such as seasonal load changes, device maintenance cycles, etc. The special structure of LSTM allows it to "remember" long-term information while "forgetting" irrelevant information, which is very suitable for handling long-period changes in energy systems. The gated recurrent unit (GRU) is used to model medium-term dependencies, such as intra-day load fluctuations, energy storage charging and discharging cycles, etc. Compared with LSTM, the GRU structure is more simplified and has higher computational efficiency, and performs well on medium-complexity time series problems. For short-term rapid changes, such as frequency fluctuations, voltage dips, etc., the system adopts a temporal convolution network (TCN) that can efficiently capture local temporal patterns. In addition to these basic time series models, the system also introduces attention mechanism enhanced time series modeling, allowing the model to dynamically focus on the most relevant historical moments based on the context. For example, when predicting photovoltaic power generation, the model will automatically focus on the historical performance under similar weather conditions. Through these time series modeling techniques, the system identifies various time-dependent relationships of device operation, such as "lead-lag" relationships (a change in one device precedes another), "cause-response" relationships (a change in one device leads to a change in another), "synergistic-inhibitory" relationships (mutual enhancement or inhibition between devices), etc. Combining these time-dependent relationships with the previously obtained spatial correlations, the system forms a complete spatio-temporal correlation pattern, which lays the foundation for the overall understanding and predictive control of the multi-energy system.

[0128] Specifically, the system represents the correlation between devices as a graph structure G = (V, E), where the nodes V represent the devices and the edges E represent the correlation strength between the devices. The graph convolution network GCN is used to extract features from this graph structure, and the convolution formula is where D is the degree matrix, A is the adjacency matrix, H(l) is the node feature of the lth layer, and W(l) is the learnable weight matrix. Through GCN, the system can learn network characteristics such as load transfer path and fault propagation pattern. At the same time, the system uses LSTM or GRU network to model time series data and identify the time-dependent relationship of device operation to obtain a complete spatio-temporal correlation pattern.

[0129] S3.2.3: Perform parameter aggregation and local optimization of the second layer network on the spatio-temporal correlation pattern, use weighted averaging, update the parameters of the second layer network according to the client data volume weight and loss function gradient, and obtain the regional collaborative representation result.

[0130] After receiving and processing the feature vectors of each local client, the edge server needs to perform parameter aggregation and local optimization on the spatio-temporal correlation pattern. The system first evaluates the performance and contribution of each client model, calculates the client weight based on data quality, model accuracy, and running stability, etc. Then, the system aggregates the model parameters uploaded by each client using weighted averaging, and the client with higher weight has greater influence on the final model. After aggregation, the system performs local optimization on the edge server based on the regional verification dataset, fine-tunes the model parameters through gradient descent algorithm, and improves the generalization ability of the model in the regional range. This process pays special attention to the processing of regional heterogeneity problems, such as differences caused by different device types and different energy characteristics. The system balances the learning effect of the model on different energy types through regularization technology and multi-task learning framework, and ensures that the optimized model can adapt to the diversified goals such as high precision demand of power system and high stability demand of heat system.

[0131] Based on the results of parameter aggregation and local optimization, the system generates an updated regional model and distributes the updated parameters back to each local client. This update process uses an incremental learning strategy, only transmitting the parameter change part, which greatly reduces the communication overhead. In order to adapt to the dynamic characteristics of energy systems, the system sets an adaptive update frequency mechanism, increases the update frequency during periods of severe energy load fluctuations (such as extreme weather conditions), and reduces the update frequency during stable system operation, balancing learning effect and computing resource consumption. At the same time, the system also implements a model rollback mechanism, which can quickly fall back to the previous stable version when detecting that the performance of the updated model has decreased, ensuring the safety of system operation. This flexible regional model update strategy enables the system to maintain learning continuity while responding to various complex operating scenarios, continuously improving the intelligent scheduling level of regional energy systems.

[0132] Specifically, the edge server performs parameter aggregation and local optimization on the spatio-temporal correlation pattern. The edge server j aggregates the gradients of the clients in the region using weighted averaging to update the regional model parameters where η is the learning rate, Cj is the set of clients under the jurisdiction of edge server j, ki is the data volume weight of client i, is the normalized weight coefficient, indicating the weight proportion of client i in parameter aggregation, L i is the loss function gradient of client i, indicating the gradient information of the client's data under the current model parameters. Through local optimization, the edge server can generate a representative result with regional collaborative cognitive ability.

[0133] Exemplarily, the parameter aggregation and local optimization of the spatio-temporal correlation pattern by the second layer network is a key step to realize the regional collaborative representation. In the hierarchical learning architecture of the multi-energy system, the second layer network is deployed on the edge server, responsible for the collaborative analysis and optimization within the region. Parameter aggregation refers to the merging of model parameters of multiple devices or subsystems within the region to form a unified regional model. The main challenge faced in this process is how to handle the heterogeneity of different device models and the difference in data distribution. To solve this problem, the system adopts a weighted average method for parameter aggregation. Unlike simple arithmetic average, the weighted average assigns different weights to the parameters of different devices according to certain criteria, making the aggregation result more accurately reflect the overall characteristics of the region. The system considers two main weight factors: client data volume weight and loss function gradient. The client data volume weight is determined based on the amount of training data provided by the device, and the device with more data gets higher weight. This design is based on the assumption that more training data usually means more reliable model parameters. However, considering only data volume may lead to bias, as data quality is also important. Therefore, the system introduces a weight adjustment mechanism based on the loss function gradient. Specifically, the system evaluates the contribution of each device model parameter to the regional objective function, and the device with larger parameter gradient (i.e., more significant impact on the objective function) gets higher weight. This dual weight mechanism ensures that the aggregation process considers both data basis and model performance. On the basis of parameter aggregation, the system also performs local optimization, i.e., further adjusts the aggregated model parameters using regional data. This step usually uses gradient descent or its variants, fine-tuning the model parameters for region-specific goals such as load balancing, operating cost, system reliability, etc. An important innovation of local optimization is the use of "progressive regularization" technology, i.e., gradually increasing the preservation constraint of global knowledge during the optimization process, avoiding overfitting to local characteristics and losing general knowledge. Through this combination strategy of parameter aggregation and local optimization, the system finally obtains the regional collaborative representation result, which not only retains the key characteristics of each device, but also captures the collaborative relationship between devices, providing a unified data basis for multi-energy coordination within the region.

[0134] Weighted averaging is an important mathematical operation that plays a central role in parameter aggregation for multi-agent systems. Unlike ordinary averaging, where all values are given equal weight, weighted averaging assigns different weights to each value based on its importance or relevance. In the process of regional collaborative representation, weighted averaging is used to integrate model parameters from different devices or subsystems. The amount of client data in the system is the first consideration for weight, which is based on the principle that devices with more training data can generally produce more reliable model parameters. For example, if there are three devices in a region, each with 1000, 2000, and 3000 training data, their parameter weights may be set to 1 / 6, 2 / 6, and 3 / 6 in a simplified case. This ensures that devices with a more solid data foundation have more say in the final model. However, data volume is not always a guarantee of quality, so the system introduces weight adjustment based on loss function gradients. The loss function is a mathematical expression that measures the difference between model predictions and actual values, while the gradient represents the rate of change of the loss function with respect to model parameters. The gradient size reflects the degree of influence of the parameter on model performance: the larger the gradient, the greater the potential contribution of parameter adjustment to error reduction. In weight allocation, the system evaluates the contribution of each device model parameter to the regional target, and parameters with more significant gradients receive higher weights. This design makes the model pay more attention to parameter updates that can effectively improve overall performance. In actual implementation, the system adopts a dynamic weight adjustment mechanism, where weights are not fixed statically but are dynamically updated as the training process and data characteristics change. For example, the system may periodically evaluate the performance of each device model on validation data and adjust weights accordingly. In addition, the system also implements an anomaly detection mechanism that automatically reduces the weight of a device when its parameter updates are abnormal (possibly due to device failure or data pollution), protecting the quality of the aggregation result. Through this carefully designed weighted averaging strategy, the system can effectively integrate the knowledge of different devices and form accurate and robust regional collaborative representation results.

[0135] S3.3: In the third layer network deployed in the cloud server, the federated averaging algorithm is used to aggregate and optimize the regional collaborative representation results, and the cross-regional coordination mode is obtained through split learning, to obtain the global intelligent model with multi-device collaborative cognitive ability.

[0136] S3.3 is the global model aggregation and strategy generation stage of the cloud server layer. The cloud server is located at the top of the entire system and is responsible for integrating the learning achievements from different regions to form a global intelligent model. The system deploys a deep neural network structure in the cloud and receives regional representation results uploaded by edge servers. These results contain the operation mode and optimization strategy of different regional energy systems, reflecting the energy characteristics and scheduling needs of each region. The cloud server uses a federated averaging algorithm to optimize the global model, which considers factors such as data size, energy type distribution, and historical performance of each region to allocate appropriate weights to different regions. During the aggregation process, the system pays special attention to the energy complementarity and synergy across regions, such as the possibility of complementary scheduling between regions rich in wind power resources and regions rich in photovoltaic resources. Based on the aggregated global model, the cloud server generates intelligent scheduling strategies for the entire system, including cross-regional energy scheduling plans, emergency response schemes, and long-term energy planning suggestions. These strategies are distributed to edge servers and local clients through a hierarchical distribution mechanism to guide the coordinated operation of energy systems at different levels. Through this bottom-up feature learning and top-down strategy distribution, the system realizes the coordinated improvement from local device cognition to global system intelligence while protecting data privacy, forming a global intelligent model with multi-device collaborative cognition ability.

[0137] Specifically, the cloud server deploys a 6-10 layer deep network to receive the representation results from each edge server. The cloud uses a federated averaging algorithm to aggregate the global model, with the update formula w = ∑mj=1αjwj, where m is the number of edge servers, αj is the weight coefficient of each edge server, usually determined according to factors such as the number of devices in the region, data quality, network delay, etc. Through global aggregation, the cloud server can learn coordination patterns, network optimization strategies, emergency response mechanisms, and other high-level cognitive abilities across regions, forming a global intelligent model with multi-device collaborative cognition ability.

[0138] Optionally, federated averaging algorithm is used to aggregate and optimize the regional collaborative representation results, which is the core mechanism to realize global collaboration. Federated averaging (FedAvg) is one of the most basic and important algorithms in federated learning, which allows multiple participants to collaboratively train a global model without sharing raw data. In a multi-energy system, different regional energy systems may be managed by different operating entities, and there is a need for data privacy protection. Federated averaging algorithm exactly solves this challenge. Its working principle is as follows: first, the cloud server initializes a global model and distributes it to the edge servers of each region; then, each region uses local data to train the model and obtains updated model parameters; then, each region uploads the model parameters (not the original data) to the cloud; finally, the cloud server performs weighted averaging on these parameters to form a new global model. This process is repeated to gradually improve the performance of the global model. In practical applications, the system has made several improvements to the standard federated averaging algorithm: first, an adaptive weight allocation mechanism based on model performance and data quality is introduced, which allows regions with more complete training and higher data quality to have more influence in the aggregation process; second, a hierarchical aggregation strategy is designed, which first aggregates locally among similar regions and then globally, reducing the negative impact of data distribution differences between regions; third, an asynchronous aggregation mechanism is implemented, which allows different regions to participate in global updates at different frequencies, improving the flexibility and robustness of the system. Through these optimizations, federated averaging algorithm can effectively integrate knowledge from different regions to form an intelligent model with a global perspective, while protecting the data privacy and autonomy of each region.

[0139] Split learning is a distributed learning method that vertically divides the layers of a neural network. Unlike federated learning, which trains a complete model on the client side, split learning cuts a deep neural network into multiple segments and deploys them on different computing nodes. In the cross-regional coordination of multi-energy systems, split learning is usually implemented as follows: shallow networks are deployed on edge servers to process raw data and extract preliminary features; intermediate layer networks are deployed on regional servers to integrate regional information and form regional representations; and deep networks are deployed on cloud servers to make cross-regional coordination decisions. Data only needs to be propagated forward to the split point, and then the intermediate activation values (rather than raw data) are passed on. During backpropagation, only gradient information is passed on. This design greatly reduces communication volume while enhancing data privacy protection. In the cross-regional coordination mode, split learning places particular emphasis on capturing inter-regional dependencies and coordination opportunities. For example, a region's excess renewable energy can provide low-cost power to neighboring regions; and a region's combined heat and power unit adjustment can affect the flow of natural gas in adjacent regions. To effectively capture these cross-regional coordination patterns, the system deploys a special "coordination layer" on the cloud, using attention mechanisms and graph neural networks to learn about resource complementarity and constraint transmission between regions. In addition, the system implements an "adaptive split point" technology that can dynamically adjust the network's split location based on computing resource distribution and network conditions, optimizing computing and communication efficiency. Through this split learning approach, the system can identify and utilize cross-regional coordination opportunities while protecting regional autonomy, ultimately forming a global intelligent model with multi-device collaborative cognitive capabilities, providing decision support for the overall optimization of multi-energy systems.

[0140] In this embodiment, a neural architecture search network with hybrid bionic learning rules is constructed in the global intelligent model, different scheduling targets such as second-level frequency stability, minute-level load suppression, and hour-level economic optimization are set up, and optimal bionic learning rule combination search is performed in the neural architecture search network to generate and execute a hierarchical intelligent scheduling strategy, including:

[0141] S4.1: In the global intelligent model, a reinforcement learning controller based on a long short-term memory network is used to construct the neural architecture search framework, and the current network architecture, historical performance indicators, and system operating conditions are used as the state space to obtain a neural network architecture search state representation;

[0142] Building a neural architecture search framework in a global intelligent model is a key step to achieve an adaptive scheduling strategy. The system uses a reinforcement learning controller based on a long short-term memory network as the core architecture, which fully considers the sequence and time-dependent characteristics of integrated energy system scheduling decisions. The long short-term memory network has a unique gating mechanism that can effectively process long sequence data and maintain long-term memory of important information, making it particularly suitable for energy system applications with complex timing characteristics.

[0143] The construction of the state space covers information in three key dimensions. First, the encoding representation of the current network architecture, the system converts the architecture features such as the number of layers, connection mode, and activation function type of the neural network into a numerical vector, forming a digital description of the architecture. Second, the historical performance indicators, including the historical records of scheduling accuracy, response speed, energy efficiency, and other key performance parameters, reflect the performance of different architectures in actual applications. Finally, the system operating condition information, including the current load level, renewable energy output status, and equipment operating state, reflects the current scheduling challenges and constraints faced by the system.

[0144] By encoding and processing these multi-dimensional state information through a long short-term memory network, the system can capture the complex correlation between state variables and generate high-quality state representations. This state representation not only contains the current system information, but also integrates historical experience and trend prediction, providing a rich information base for subsequent architecture search decisions.

[0145] Specifically, the system uses a reinforcement learning controller based on a long short-term memory network (LSTM) to build a neural architecture search framework. The state space includes the encoding representation of the current network architecture, the performance indicators (such as scheduling accuracy, response speed, energy efficiency, etc.) during the historical training process, and the current operating conditions of the system (such as load level, renewable energy output, equipment state, etc.). The controller encodes these state information through an LSTM network to obtain the state representation of neural network architecture search.

[0146] S4.2: Based on the neural network architecture search state representation, use the reinforcement learning controller to sample candidate architectures, generate candidate neural network architectures through the action space of layer selection, activation function type, connection mode, and learning rule combination, and obtain a diverse set of candidate neural network architectures;

[0147] The candidate architecture sampling process is the core link of neural architecture search, and the system intelligently generates architectures based on the output of the reinforcement learning controller. This process is not random, but based on a deep analysis of historical performance and current state, it strategically generates architecture combinations that are most likely to achieve excellent performance.

[0148] The design of the action space covers all key elements of the neural network architecture. In terms of the number of layers, the system sets corresponding ranges of the number of layers for different biomimetic learning rules, for example, a spiking neural network usually adopts a shallow structure of two to five layers to ensure fast response, a Hebbian learning network adopts a medium-depth structure of three to six layers to balance learning ability and stability, and a backpropagation network can adopt a deep structure of four to eight layers to obtain stronger expression ability.

[0149] The selection of the activation function type has an important influence on the network performance, and the system provides multiple choices including the rectified linear unit, the hyperbolic tangent function, the sigmoid function, etc., each of which has its specific applicable scenario and advantage. The design of the connection mode includes full connection, residual connection, dense connection, etc., and these different connection modes can affect the propagation path and calculation efficiency of information in the network.

[0150] The combination of learning rules is the core innovation of this patent, and the system organically combines the time coding of spiking neural networks, the adaptive mechanism of Hebbian learning, the feature selection of competitive learning, and the global optimization of backpropagation, etc. multiple biomimetic learning rules to form a hybrid learning mechanism that adapts to different scheduling goals.

[0151] The controller can generate a large number of diverse candidate network architecture sets by intelligently sampling different options in these action spaces. This diversity ensures that the search process can cover a wide solution space and avoid falling into a local optimal solution.

[0152] Specifically, the controller samples candidate architectures based on state representations, and the action space includes network layer selection (such as 2-5 layer SNN, 3-6 layer Hebbian network, 4-8 layer BP network), activation function type (ReLU, Sigmoid, Tanh, etc.), connection mode (full connection, residual connection, dense connection, etc.), and learning rule combination (SNN time coding, Hebbian learning, competitive learning, BP learning, etc.). The controller generates a diverse set of candidate network architectures by sampling options in these action spaces.

[0153] S4.3: Perform neural network architecture performance evaluation on the diverse set of candidate neural network architectures on the historical running data validation set, calculate the reward signal through the comprehensive index of scheduling accuracy, response speed, and energy efficiency, and obtain the neural network architecture performance evaluation result;

[0154] Performance evaluation of neural network architectures is a crucial step in ensuring the quality of the search. The system uses a validation method based on historical operation data to evaluate the actual performance of each candidate architecture. The validation dataset usually contains at least a year of complete operation data, covering various typical operation scenarios, extreme working conditions and seasonal changes, ensuring the comprehensiveness and reliability of the evaluation results.

[0155] The scheduling accuracy evaluation mainly focuses on the accuracy of the architecture in prediction and control, including load prediction error, frequency control accuracy, temperature regulation deviation and other key indicators. The system compares the difference between the scheduling instructions output by the architecture and the actual optimal solution to quantify the scheduling accuracy of the architecture. Response speed evaluation focuses on the reaction ability of the architecture in the face of sudden events or rapid changes, including scheduling strategy generation time, control instruction transmission delay, system state adjustment speed and other time-related indicators.

[0156] Energy efficiency evaluation covers economic and environmental indicators such as energy utilization rate, operating cost, carbon emission level, etc. These indicators reflect the control ability of the architecture in achieving scheduling goals while controlling resource consumption and environmental impact. The system also introduces a network complexity penalty term to avoid unnecessary computational burden and maintenance difficulties caused by overly complex architectures.

[0157] Through the comprehensive evaluation of these multi-dimensional indicators, the system can calculate the reward signal reflecting the overall performance of the architecture. This reward signal not only considers the pros and cons of a single indicator, but also balances the trade-off between different indicators, providing high-quality feedback information for the learning of the controller.

[0158] Specifically, the system will train and evaluate the generated candidate architectures on a validation set containing historical 1-year operation data. Evaluation metrics include scheduling accuracy (prediction error, control accuracy, etc.), response speed (scheduling strategy generation time, control instruction execution delay, etc.), energy efficiency (energy utilization rate, economic cost, etc.). The system calculates the reward signal R = w1·Accuracy + w2·Speed + w3·Efficiency - w4·Complexity through the weighted combination of these indicators, where wi (w1, w2, w3, w4) is the weight coefficient, and Complexity is the network complexity penalty term. Among them, Accuracy is the adjustment accuracy, reflecting the accuracy of the architecture in prediction and control, which includes: load prediction error, frequency control accuracy, temperature regulation deviation, etc. Key indicators can be quantitatively evaluated by comparing the differences between the scheduling instructions output by the architecture and the actual optimal solution; Speed represents the response speed, focusing on the reaction ability of the architecture in the face of sudden events or rapid changes, including: scheduling strategy generation time, control instruction transmission delay, system state adjustment speed, etc. Time-related indicators are used to evaluate the real-time response ability of the system; Efficiency represents energy efficiency, which covers energy utilization rate, operating cost, carbon emission level, etc. Economic and environmental indicators reflect the control ability of the architecture in achieving scheduling goals while controlling resource consumption and environmental impact.

[0159] S4.4: According to the performance evaluation results of the neural network architecture, the policy gradient algorithm is used to optimize the controller strategy parameters, and the different scheduling targets of the second-level frequency stability, minute-level load suppression, and hour-level economic optimization are used to guide the architecture search direction, and the optimal bionic learning rule combination is obtained;

[0160] The application of the policy gradient algorithm is the core mechanism of the reinforcement learning controller optimization. The system continuously iterates and improves the decision-making strategy of the controller, enabling it to generate increasingly superior network architectures. This optimization process is based on the basic principles of reinforcement learning, that is, obtaining feedback through interaction with the environment and adjusting the behavior strategy according to the feedback.

[0161] The key advantage of the policy gradient algorithm is its ability to directly optimize expected returns without the need for explicit modeling of the value function. In the application of neural architecture search, this means that the controller can directly learn how to generate high-performance architectures without the need to predefine the value assessment of all possible architectures. The algorithm calculates the gradient of the policy parameters to determine the direction and magnitude of parameter adjustment, gradually improving the decision-making ability of the controller.

[0162] The guidance mechanism of different scheduling objectives is an important innovation of the system. The second-level frequency stability objective requires extremely fast response speed and high-precision control ability, so the architecture search will tend to select network combinations with fast activation characteristics and simple structure. The minute-level load suppression objective needs to balance response speed and prediction accuracy, and the architecture search will look for network structures that can effectively handle medium-term changes. The hour-level economic optimization objective is more concerned about global optimality and long-term stability, and the search process will tend to select deep network structures with strong expression and optimization capabilities.

[0163] Through this target-oriented search mechanism, the system can ultimately obtain optimal bionic learning rule combinations for different time scale scheduling requirements. These combinations are not fixed, but are dynamically adjusted according to changes in system operating conditions and external environment, ensuring that the scheduling strategy always remains optimal.

[0164] Specifically, the system uses the REINFORCE policy gradient algorithm to optimize the controller parameters based on the performance evaluation results. The gradient calculation formula is where is the expected operator, representing the expectation of all possible trajectories under the policy πθ, πθ is the controller policy, a is the sampled architecture action, s is the state representation, and R is the reward signal. Through continuous iteration, the controller gradually learns the policy that can generate high-performance architectures, and ultimately obtains optimal bionic learning rule combinations suitable for different scheduling objectives.

[0165] S4.5: For the optimal bionic learning rule combination, route different frequency components to the corresponding bionic learning layer through the time scale decomposer, use the attention mechanism for multi-layer information fusion, and generate and execute the hierarchical intelligent scheduling strategy.

[0166] The generation and execution of the hierarchical intelligent scheduling strategy is the ultimate goal of the entire system, and this stage converts the results of the previous steps into actual control actions. The time scale decomposer is crucial as it can accurately identify different frequency components in the scheduling requirements and route them to the corresponding bionic learning layer for specialized processing.

[0167] The high-frequency components mainly correspond to control tasks that require fast response, such as grid frequency stability. These components are routed to the spiking neural network layer. Spiking neural networks have event-driven computing characteristics and can respond to input changes within milliseconds, making them ideal for handling fast control tasks such as frequency regulation. The medium-frequency components correspond to tasks on a medium time scale, such as load fluctuation regulation, and are routed to the Hebbian learning layer. Hebbian learning rules can adaptively adjust neuron connection strengths, making them particularly suitable for handling load regulation tasks with pattern changes. The low-frequency components correspond to long-term planning tasks such as economic optimization and are routed to the backpropagation layer. Backpropagation algorithms have strong global optimization capabilities and can handle complex multi-objective optimization problems.

[0168] Information fusion between layers is achieved through an attention mechanism, which can dynamically assess the importance of different layer outputs and assign weights accordingly. The introduction of the attention mechanism enables the system to automatically adjust the priority of different time scale tasks based on the current situation, achieving truly intelligent scheduling.

[0169] The final generated layered intelligent scheduling strategy is a comprehensive control scheme that considers different needs on the second, minute, and hour levels, achieving overall optimal operation of the system through coordination. This layered architecture not only improves the accuracy and response speed of scheduling, but also enhances the system's adaptability to uncertainties and disturbances.

[0170] Specifically, the system routes scheduling requirements of different frequency components to the corresponding biomimetic learning layers through a time scale decomposer. High-frequency components (>1Hz) are routed to the SNN layer for fast response, medium-frequency components (0.1-1Hz) are handled by the Hebbian learning layer for load fluctuations, and low-frequency components (<0.1Hz) are handed over to the BP layer for economic optimization. Information fusion between layers is achieved through an attention mechanism, with a weight distribution formula of αi = exp(ei) / ∑jexp(ej), where ei = tanh(Wi·hi + bi) is the importance score of the i-th layer. Through this layered architecture, the system can generate and execute intelligent scheduling strategies that take into account the needs of different time scales.

[0171] In this embodiment, the layered intelligent scheduling strategy includes a second-level frequency stability control strategy:

[0172] S4.5.1: Identify and extract the first frequency component in the layered intelligent scheduling strategy, and convert the frequency deviation signal into a pulse firing rate proportional to the deviation amplitude through a Poisson encoder to obtain a frequency deviation pulse sequence;

[0173] The identification and extraction of frequency deviation signals is a fundamental step in the control of frequency stability at the second level. This process requires the accurate separation of high-frequency components related to frequency stability from complex system signals. Through advanced signal processing techniques, the system can monitor the small changes in the grid frequency in real time and quickly identify the frequency deviation signals that require immediate response. This identification process not only considers the magnitude of the deviation but also analyzes the trend and duration of the deviation to ensure that the control system can accurately determine the urgency of frequency stability.

[0174] The application of the Poisson encoder is a key technology for converting continuous frequency deviation signals into discrete pulse sequences. This encoding method draws on the information processing mechanism of the biological nervous system, in which neurons encode information intensity through the firing rate of pulses. In frequency control applications, when the frequency deviation is larger, the Poisson encoder generates a higher pulse firing rate, forming a natural intensity encoding mechanism. The advantage of this encoding method is that it can maintain the temporal dynamic characteristics of the signal while converting continuous signals into discrete pulse forms that are more suitable for processing by spiking neural networks.

[0175] The parameter settings in the encoding process have a significant impact on the control effect. The encoding gain parameter determines the mapping relationship between the deviation signal and the pulse firing rate. A higher gain makes the system more sensitive to small amplitude deviations, but it may also increase the influence of noise. The bias parameter is used to set the dead zone range of the encoding, avoiding unnecessary responses of the system to small measurement noise. By carefully adjusting these parameters, the system can maintain high sensitivity while avoiding excessive responses.

[0176] The generated frequency deviation pulse sequence not only contains the amplitude information of the deviation but also retains the time characteristics of the deviation change. This time encoding characteristic enables the subsequent spiking neural network to better understand the dynamic process of frequency change, thereby generating more accurate and timely control responses.

[0177] Specifically, the system identifies and extracts the first frequency in the scheduling strategy, i.e., the high-frequency component (>1 Hz), mainly for grid frequency stability control. The system converts the measured frequency deviation signal Δf through the Poisson encoder into a pulse firing rate r(t) = λ·max(0, Δf + b), where λ is the encoding gain and b is the bias parameter. When the frequency deviation is larger, the generated pulse sequence density is higher, achieving the mapping of signal intensity to time encoding.

[0178] S4.5.2: Based on the frequency deviation pulse sequence, a leaky integrate-and-fire model of the spiking neural network is used for membrane potential dynamic modeling, and the membrane potential change is calculated through the time constant, resting potential, and input resistance of the membrane potential, obtaining a pulse response signal;

[0179] The leaky integrate-and-fire model is the core computational unit of spiking neural networks, which accurately simulates the electrophysiological properties of biological neurons. In frequency control applications, each artificial neuron processes the incoming pulse sequence through this model and decides whether to generate an output pulse based on the accumulated stimulus intensity. This computational mechanism is fundamentally different from traditional artificial neural networks, which are simple weighted summation and activation function calculations, but a continuous dynamic process.

[0180] The dynamic modeling process of membrane potential involves the interaction of multiple key parameters. The membrane time constant determines the response speed and memory duration of the neuron to input stimuli, a smaller time constant enables the neuron to respond quickly to input changes, suitable for processing high-frequency signals, while a larger time constant makes the neuron have longer memory ability, suitable for accumulating long-term information. Resting potential represents the basic state of the neuron without external stimulation, which affects the activation threshold and recovery characteristics of the neuron. The input resistance determines the degree of influence of external current on membrane potential, a higher input resistance makes the neuron more sensitive to weak signals.

[0181] In the specific application of frequency control, the setting of these parameters needs to carefully balance the requirements of response speed and stability. For frequency stability control, the system needs to respond to frequency changes within milliseconds, so the membrane time constant is usually set small to ensure fast response. At the same time, in order to avoid excessive sensitivity to measurement noise, the resting potential and input resistance need to be set to provide appropriate filtering effect.

[0182] The calculation process of membrane potential is a continuous differential equation solving process, the system updates the membrane potential state of each neuron in real time through numerical integration method. This continuous calculation enables the neuron to smoothly process the changes of input signal, avoiding the sudden phenomena that may occur in traditional discrete calculation. The generated pulse response signal not only reflects the current input intensity, but also integrates the influence of historical information, providing a more stable and reliable basis for frequency control.

[0183] Specifically, the system uses the Leaky Integrate-and-Fire (LIF) model of spiking neural networks (SNN) to process the frequency deviation pulse sequence. The membrane potential dynamic equation of the LIF model is τm·dV / dt = -(V-Vrest) + R·I(t), where τm is the membrane time constant, V is the membrane potential, Vrest is the resting potential, R is the input resistance, and I(t) is the input current (generated by input pulses). By solving this equation, the system obtains the change of neuron membrane potential with time, forming a response signal to the frequency deviation.

[0184] S4.5.3: Thresholding the pulse response signal, generating a control pulse and resetting the membrane potential when the membrane potential exceeds a firing threshold, obtaining a frequency regulation instruction;

[0185] The thresholding process of the pulse response signal is the key mechanism for the spiking neural network to generate output. This process simulates the mechanism of action potential generation of biological neurons. When the membrane potential of a neuron reaches a certain threshold, a standardized output pulse is generated, and then the membrane potential is reset to a lower level, preparing for the next activation. This all-or-nothing response characteristic makes the spiking neural network have natural digital processing capability, while maintaining the time dynamic characteristics of biological systems.

[0186] The setting of the firing threshold has a key impact on the performance of the control system. A lower threshold makes the neuron more easily activated, improving the sensitivity of the system, but it may also lead to excessive response and instability. A higher threshold improves the stability of the system, but may reduce the response ability to small amplitude signals. In frequency control applications, the setting of the threshold needs to be optimized according to the specific characteristics of the power grid and the control requirements, ensuring sufficient response sensitivity while maintaining stability.

[0187] The membrane potential reset mechanism is an important link to maintain the normal work of neurons. After the neuron fires a pulse, the membrane potential is not simply returned to the resting potential, but is set to a specific reset potential, which is usually lower than the resting potential. This design simulates the refractory period characteristics of biological neurons, preventing neurons from firing continuously in a short period of time, and helps to improve the stability and controllability of the output signal.

[0188] The conversion process from pulse output to specific control instruction involves complex signal processing and decision logic. The system converts the pulse output of neurons into specific control signal strength through the synaptic weight matrix, which represents the influence degree of different neurons on different control objects. The learning and adjustment of weights are important aspects of system performance optimization. Through continuous training and adjustment, the system can learn the optimal control strategy.

[0189] The generated frequency regulation instruction includes the setting value adjustment of the generator set governor and the power instruction of the energy storage system. These instructions not only contain the direction and amplitude of control, but also contain the time requirements and priority information of execution. The generation process of the instruction considers various constraints such as physical limitations of devices, response speed and adjustment ability, ensuring that the generated instruction can effectively improve the frequency deviation without exceeding the safe operating range of the device.

[0190] Specifically, the system performs threshold judgment on the membrane potential of neurons, and when V > Vth (Vth is the firing threshold), the neuron generates an output pulse and resets the membrane potential to Vreset. The output pulse is converted into a specific control signal strength through a synaptic weight matrix to generate frequency regulation instructions, such as generator set speed governor set value adjustment, energy storage system power instruction, etc.

[0191] S4.5.4: According to the frequency regulation instruction, millisecond-level power regulation is performed on the generator set and energy storage device, and the frequency deviation is eliminated through primary frequency control;

[0192] The execution process of the control instruction is the key link for the frequency stability control system to play an actual role. This process needs to coordinate various types of power generation and energy storage equipment to achieve rapid and accurate regulation of the grid frequency. The system transmits frequency regulation instructions to each execution device through a high-speed communication network to ensure that the instructions can reach the target device and start execution in the shortest time.

[0193] The speed regulation system of the generator set is the main means of traditional grid frequency control. These systems change the mechanical power output of the generator by adjusting the steam intake of the steam turbine or the guide vane opening of the water turbine. Modern speed regulation systems have millisecond-level response capability and can quickly adjust the output power of the generator set after receiving the control instruction. Different types of generator sets have different regulation characteristics. Although thermal power generators have large regulation amplitude, their response is relatively slow. The response of hydroelectric generators is fast, but their regulation capacity is affected by water level. Gas turbines have fast start-stop and regulation capabilities, but the operating cost is relatively high.

[0194] The power control system of the energy storage device is an important supplement to modern grid frequency regulation, especially the battery energy storage system, which has extremely fast response speed and precise power control capability. The energy storage system can switch from charging to discharging state within milliseconds, or quickly adjust the size of charging and discharging power. This fast response capability makes the energy storage system particularly suitable for handling rapid fluctuations and transient imbalances of frequency.

[0195] The implementation of millisecond-level power regulation requires precise control technology and high-performance actuators. The system uses advanced power electronics technology to achieve precise regulation of power output through high-frequency switching and precise pulse width modulation control. At the same time, the system also uses predictive control technology, which can start adjusting device output in the early stage of detecting frequency deviation, further shortening the overall response time.

[0196] Primary frequency control is the first line of defense for grid frequency stability. It activates within seconds of a frequency deviation through automatic responses from devices, effectively preventing further deterioration of frequency. This control mechanism does not require external communication or complex calculations, but rather direct feedback control based on local frequency measurements, with high reliability and fast response. Through the rapid intervention of primary frequency control, the system can control the frequency deviation within an acceptable range before it expands, giving time for subsequent secondary frequency control and economic dispatch.

[0197] Specifically, the system issues frequency regulation instructions to each execution device through a communication network, including the speed control system of the generator set and the power control system of the energy storage device. These devices adjust active power output within milliseconds according to the instructions, achieving fast response and primary frequency control of the grid frequency, eliminating frequency deviation.

[0198] S4.5.5: After eliminating the frequency deviation, real-time monitoring and feedback of the frequency control state are performed to control the frequency stability error within the range of plus or minus 0.05 Hz.

[0199] Real-time monitoring and feedback of the frequency control state is a key mechanism to ensure the continuous and effective operation of the frequency stability control system. The system monitors the changes in grid frequency in real time through high-precision frequency measurement devices distributed at key nodes in the grid. These measurement devices use advanced digital signal processing technology to accurately extract frequency information in a noisy environment and provide millisecond-level measurement update frequency.

[0200] The calculation and evaluation of frequency stability error is the core function of the monitoring system. The system not only calculates the instantaneous deviation between the current frequency and the standard frequency, but also analyzes the trend and speed of frequency change to predict possible development directions. This comprehensive analysis allows the system to distinguish between temporary small fluctuations and persistent deviations that may lead to system instability, enabling more intelligent control decisions.

[0201] Dead zone setting is an important mechanism for the frequency control system to avoid overaction. When the frequency deviation fluctuates within a small range, the system will not initiate control response, which can avoid unnecessary control action on normal small fluctuations and reduce device wear and energy loss. The size of the dead zone needs to balance the requirements of control accuracy and system stability, and is usually determined according to the specific characteristics of the grid and operational experience.

[0202] The implementation of closed-loop control mechanism ensures the continuity and adaptability of frequency control. When the frequency deviation is detected to be outside the dead zone range, the system will re-evaluate the current situation and generate new control instructions. This continuous monitoring and adjustment process allows the control system to adapt to changes in grid operating conditions, maintaining long-term stability of control effectiveness.

[0203] The control accuracy requirement of ±0.05 Hz reflects the strict standard of frequency stability in modern power grids. This accuracy requirement not only ensures the safe and stable operation of the power grid, but also ensures that various devices connected to the power grid can work normally. Excessive deviation of frequency can lead to serious consequences such as device damage, protection device malfunction or system collapse. Through accurate frequency control, the system can maintain high-quality operation of the power grid, ensuring the reliability and safety of power supply.

[0204] The maintenance of control accuracy requires the coordinated cooperation of various links of the control system, from signal detection, instruction generation to execution response, each link needs to meet the corresponding performance requirements. The system ensures that the control accuracy can be maintained in the required range for a long time through continuous performance monitoring and parameter optimization. When the control performance is detected to be decreased, the system will automatically adjust the control parameters or switch the control strategy to ensure the reliability and effectiveness of frequency control.

[0205] Specifically, the system continuously monitors the frequency state of the power grid and calculates the frequency control error e = f - fref, where f is the actual frequency and fref is the nominal frequency (such as 50 Hz). When the error exceeds the preset dead zone (such as ±0.01 Hz), the frequency control process is retriggered. Through this closed-loop control mechanism, the system can always control the frequency stability error within ±0.05 Hz, ensuring the stable operation of the power grid.

[0206] In this embodiment, the hierarchical intelligent scheduling strategy includes a minute-level load leveling control strategy:

[0207] S4.5.6: Extract the second frequency component in the hierarchical intelligent scheduling strategy, enhance the adaptive ability of load prediction using Hebb learning rule, and obtain an adaptive load prediction model by simultaneously activating the enhanced learning load change pattern of neuron connection strength;

[0208] The Hebbian learning rule is a fundamental theory in the fields of neuroscience and machine learning, first proposed by Canadian psychologist Donald Hebb in his book "The Organization of Behavior" in 1949. This rule is succinctly summarized as "neurons that fire together, wire together," and is often expressed as "neurons that fire together, connect together." The Hebbian learning rule describes the mechanism by which neurons in the brain form connections through experience, and is the foundation of self-organizing learning in biological neural networks. Its core idea is that when two neurons are activated simultaneously, the strength of the synaptic connection between them increases; conversely, if their activities are unrelated or negatively correlated, the connection strength may decrease. This mechanism enables neural networks to automatically learn and adapt through repeated exposure to similar patterns of stimulation. In multi-energy systems, the Hebbian learning rule is innovatively applied to the learning of correlations between energy devices. For example, when the system observes that an increase in the power output of a certain cogeneration unit is often accompanied by the startup of a specific thermal load device, it automatically strengthens the correlation weight between the two devices, forming an "energy neural network." This learning method does not require explicit supervision signals, but rather discovers intrinsic correlations by observing patterns of coordinated activation of devices, making it well-suited to capture complex and possibly not fully understood by human experts device interaction patterns in multi-energy systems. As the system accumulates operational data, the Hebbian rule-based learning mechanism can gradually establish a connection network that reflects the true correlation strength between devices, providing important basis for system understanding and decision-making. In practical applications, the Hebbian learning rule is often combined with other learning algorithms to form a more powerful hybrid learning system, retaining the advantages of self-organizing learning of the Hebbian rule while overcoming its slow convergence and potential instability in certain scenarios.

[0209] Real-time monitoring and feedback of frequency control status is a key mechanism to ensure the continuous and effective operation of the frequency stability control system. The system monitors the changes in grid frequency in real time through high-precision frequency measurement devices distributed at key nodes in the grid. These measurement devices use advanced digital signal processing technology to accurately extract frequency information in a noisy environment and provide millisecond-level measurement update frequency.

[0210] Calculation and evaluation of frequency stability error is the core function of the monitoring system. The system not only calculates the instantaneous deviation between the current frequency and the standard frequency, but also analyzes the trend and speed of frequency change to predict possible development directions. This comprehensive analysis enables the system to distinguish between temporary small fluctuations and persistent deviations that may lead to system instability, making more intelligent control decisions.

[0211] The deadband setting is an important mechanism for the frequency control system to avoid overaction. When the frequency deviation fluctuates within a small range, the system will not initiate control response, which can avoid unnecessary control action on normal minor fluctuations, reduce equipment wear and energy loss. The size of the deadband needs to balance the requirements of control accuracy and system stability, usually determined according to the specific characteristics of the power grid and operating experience.

[0212] The implementation of closed-loop control mechanism ensures the continuity and adaptability of frequency control. When the frequency deviation is detected to exceed the deadband range, the system will re-evaluate the current situation and generate new control instructions. This continuous monitoring and adjustment process enables the control system to adapt to changes in power grid operating conditions, maintaining long-term stability of control effect.

[0213] The control accuracy requirement of plus or minus zero point five hertz reflects the strict standard of modern power grid on frequency stability. This accuracy requirement not only guarantees the safe and stable operation of the power grid, but also ensures that various devices connected to the power grid can work normally. Excessive deviation of frequency may cause equipment damage, protection device misoperation or system collapse, etc. Through accurate frequency control, the system can maintain high-quality operation of the power grid, ensuring the reliability and safety of power supply.

[0214] The maintenance of control accuracy requires the coordination of each link of the control system, from signal detection, instruction generation to execution response, each link needs to meet the corresponding performance requirements. The system ensures that the control accuracy can be maintained within the required range for a long time through continuous performance monitoring and parameter optimization. When the control performance is detected to decline, the system will automatically adjust the control parameters or switch the control strategy, ensuring the reliability and effectiveness of frequency control.

[0215] Specifically, the system extracts the second frequency in the scheduling strategy, i.e. the intermediate frequency component (0.1-0.99 Hz), mainly for load suppression control. The system uses Hebbian learning rule ("simultaneous activation of neurons enhances connection") to construct an adaptive load prediction model. The weight update rule is Δwij = η·xi·xj·(xj - θ), where xi and xj are the activation values of the previous and subsequent neurons, θ is the plasticity threshold, and η is the learning rate. Through this mechanism, the model can adaptively learn the load change pattern, enhancing the accuracy and robustness of prediction.

[0216] S4.5.7: According to the adaptive load prediction model, a time window mechanism is introduced to learn the periodic pattern and mutation characteristics of the load, the prediction network parameters are trained through historical load data, and the load prediction result is obtained;

[0217] The introduction of time window mechanism is a key technical means for the load forecasting system to capture multi-scale time characteristics. Different lengths of time windows can capture different levels of load change characteristics. Short windows focus on immediate changes and sudden events, medium windows identify trend changes and adjustment processes, and long windows reveal periodic patterns and seasonal rules. This multi-scale time analysis method enables the forecasting system to fully understand the complexity of load changes and avoid the limitations of single time scale analysis.

[0218] Short time window analysis mainly focuses on load mutation characteristics within five to fifteen minutes. These mutations are usually caused by rapid start-stop of equipment, emergency load input or removal, and load changes caused by sudden events. Short window analysis can quickly identify these mutation patterns and provide timely warning information for the system. At the same time, short window analysis can capture the instantaneous rate and acceleration information of load changes, providing data support for rapid response control.

[0219] Medium time window processing covers load change trends from one to four hours. Load changes in this time scale usually have certain continuity and predictability, reflecting the gradual process of user activity patterns. For example, the load change of office buildings on weekdays, the electricity usage rules of residential areas at different times, the production cycle of industrial enterprises, etc. Medium window analysis provides the basis for medium-term prediction and adjustment strategy formulation for the system by identifying these trend changes.

[0220] Long time window analysis extends to the time scale of days or weeks, mainly used to capture the periodic patterns and long-term rules of load. Daily periodicity is reflected in the load level difference at different times of the day, the change of load patterns between weekdays and weekends, and the load characteristics of special dates such as holidays. Weekly periodicity reflects the impact of work cycles on load, including regular changes on weekdays and load characteristics on weekends. The identification of these long-term patterns provides a basic framework for load forecasting for the system.

[0221] The training process of historical load data is a systematic machine learning process. The system uses complete historical data containing at least one year for model training to ensure that all possible seasonal changes and special situations are covered. Training data not only includes load values, but also related environmental factors such as weather conditions, date types, economic activity levels, etc. These auxiliary information helps the model better understand the external driving factors of load changes.

[0222] The optimization of prediction network parameters is achieved by minimizing the prediction error. The system employs various optimization algorithms, including gradient descent, adaptive learning rate adjustment, regularization techniques, etc., to ensure that the model not only accurately fits historical data but also has good generalization ability. The complexity control of the model is also considered in the parameter optimization process to avoid overfitting and ensure the stability of the prediction performance when facing new data.

[0223] Specifically, the system introduces a time window mechanism in the Hebbian learning model and uses a sliding window method to capture the load change characteristics at different time scales. Short windows (such as 5-15 minutes) are used to capture sudden changes, medium windows (such as 1-4 hours) are used to capture trend changes, and long windows (such as 1 day or 1 week) are used to capture periodic patterns. The system uses at least one year of historical load data to train the model parameters, and minimizes the prediction error function Optimize network weights to get preliminary load prediction results, where represents the actual load value, i.e. the true historical load data at time t, represents the prediction error, which represents the deviation between the actual value and the predicted value, represents the squared error, which takes the square of the deviation to eliminate the sign and amplify the larger error.

[0224] S4.5.8: Adopt a competitive learning mechanism for the load prediction results to select features, enhance the key feature recognition ability through lateral inhibition, and improve the load prediction accuracy to more than 95%, to obtain high-precision load prediction results;

[0225] The application of the competitive learning mechanism in load prediction optimization embodies the important principle of neuron competition and activation in biological neural networks. Under this mechanism, multiple neurons receive input information simultaneously, but only the neuron with the strongest response is selected for weight update, and the weights of other neurons remain unchanged or are inhibited. This "winner-takes-all" strategy enables the network to automatically identify and strengthen the most important feature patterns while suppressing the influence of noise and redundant information.

[0226] The feature selection process is automated and intelligent through competitive learning. Traditional feature selection methods usually require manual analysis and expert experience to determine which features are most valuable for prediction, while the competitive learning mechanism allows the network to autonomously learn and discover the most relevant feature combinations. In load prediction applications, the system may need to handle hundreds of potential input features, including historical load values, weather parameters, time features, economic indicators, etc. The competitive learning mechanism automatically filters out the key features that have the greatest impact on prediction results through competition between neurons.

[0227] Lateral inhibition is an important component of competitive learning mechanism, which simulates the mutual inhibition phenomenon between adjacent neurons in biological neural networks. When a neuron is strongly activated, it will inhibit the activity of surrounding neurons, thereby enhancing the contrast and clarity of the signal. In load forecasting systems, lateral inhibition helps to eliminate redundancy between similar features and highlight the most discriminative feature combinations. This mechanism is particularly suitable for dealing with the multicollinearity problem existing in high-dimensional data, improving the stability and interpretability of the model.

[0228] The enhancement of key feature identification ability is the result of the joint action of competitive learning and lateral inhibition. Through this mechanism, the system can automatically discover the core driving factors of load changes, such as the influence of temperature changes in a specific period on air conditioning load, the influence of weekday patterns on commercial load, and the influence of economic activity levels on industrial load. The accurate identification of these key features not only improves the prediction accuracy, but also provides a scientific basis for the development of load management and demand response strategies.

[0229] The goal of improving load prediction accuracy to more than 95% reflects the system's strict requirements for prediction quality. This level of accuracy requires the comprehensive application of multiple technical means, including high-quality training data, advanced algorithm architecture, fine parameter tuning, and effective feature engineering. High-precision prediction provides a reliable foundation for subsequent load flattening control, reducing control bias and resource waste caused by prediction errors.

[0230] The acquisition of high-precision load prediction results marks the maturity and reliability of the prediction system. These prediction results not only include numerical prediction of load, but also include uncertainty assessment and confidence interval estimation of prediction. The system provides prediction results with multiple time levels, from short-term prediction of a few minutes to medium-term prediction of a few hours, supporting different types of control decisions. The quality of prediction results is evaluated through multiple indicators, including mean absolute error, root mean square error, and prediction accuracy, ensuring that the prediction performance meets the requirements of practical applications.

[0231] Specifically, the system introduces a competitive learning mechanism to optimize the load prediction model. Competitive learning uses the "winner-takes-all" principle, where only the weights of the neuron with the highest activation value are updated, while suppressing other neurons. The formula is Δw ij =η·x i ·(x j - w ij ), where i is the winner neuron index. Through lateral inhibition, the system can automatically select the key features that have the most impact on prediction, filter noise and redundant information, and improve load prediction accuracy to more than 95%.

[0232] S4.5.9: Based on the high-precision load prediction results, calculate the load deviation and fluctuation trend, use energy storage systems and controllable loads for power compensation and peak shaving, and obtain the load leveling control strategy.

[0233] The calculation of load deviation and fluctuation trend is a key step in the formulation of load leveling control strategy. By comparing the actual load measurement value with the high-precision prediction result, the system can accurately identify the deviation of the load and the development trend. This analysis not only focuses on the deviation size at the current time, but also predicts the possible changes in the future, providing forward-looking information for the formulation of control strategy. Deviation analysis includes multiple dimensions such as deviation amplitude, duration, change rate, etc., helping the system to comprehensively evaluate the impact of load disturbance.

[0234] Energy storage systems play a key role in load leveling control, and their unique bidirectional power regulation capability makes them an ideal load regulation tool. Energy storage systems can discharge to supplement power shortage when load is higher than expected, and charge to absorb excess power when load is lower than expected. This fast charge-discharge switching capability makes energy storage systems particularly suitable for handling rapid fluctuations and short-term imbalances in load. The control strategy of energy storage systems needs to consider multiple factors such as current state of charge, charge-discharge efficiency, equipment life and economy, etc.

[0235] The coordinated participation of controllable loads provides demand-side regulation means for load leveling. Controllable loads include interruptible loads, transferable loads and adjustable loads, etc. Interruptible loads can be temporarily cut off when the system needs to provide immediate load reduction; transferable loads can adjust the electricity usage time from peak to off-peak, achieving time transfer of load; adjustable loads can adjust the power consumption level within a certain range, providing continuous load regulation capability. The coordinated use of these different types of controllable loads provides flexible and diverse regulation means for the system.

[0236] Power compensation mechanism is a direct implementation way of load leveling control. When the load deviation is detected, the system will quickly calculate the required compensation power and optimize the allocation between energy storage systems and controllable loads. The allocation of compensation power needs to consider various constraints such as available capacity of energy storage systems, response ability of controllable loads, technical limitations of equipment, etc. The system uses optimization algorithms to determine the optimal power allocation scheme, minimizing control cost and equipment loss while meeting the load leveling target.

[0237] Peak clipping strategy is an important application form of load leveling control. It reduces electricity demand during peak load period and increases electricity demand during low load period to smooth the load curve. Peak clipping operation is achieved through energy storage discharge and controllable load reduction, effectively reducing the peak load pressure of the system. Valley filling operation is achieved through energy storage charging and adjustable load increase, improving the load level of the system during the low valley period and improving the operation efficiency of the power generation equipment. This two-way adjustment mechanism not only improves the load characteristics of the system, but also creates favorable conditions for the accommodation of renewable energy.

[0238] The development of load leveling control strategy is a comprehensive optimization process that needs to balance multiple objectives and constraints. The objectives considered in the strategy development process include load deviation minimization, control cost minimization, device loss minimization, and user satisfaction maximization. The constraints include technical limitations of energy storage systems, response limitations of controllable loads, grid safety constraints, and user contract constraints. The system uses a multi-objective optimization algorithm to find the control strategy that best balances each objective while meeting all constraints. The final control strategy is a dynamic and adaptive solution that can adjust control parameters and execution methods according to real-time conditions.

[0239] Specifically, the system calculates the deviation ΔP = Pactual - Pforecast between the actual load and the predicted value and the load fluctuation trend dP / dt based on high-precision load prediction results. According to these indicators, the system develops load leveling control strategies, including energy storage system charging and discharging plans, controllable load switching strategies, etc. The energy storage system discharges during the load peak period and charges during the low valley period to achieve peak clipping and valley filling; controllable loads reduce electricity consumption when system pressure is high and increase electricity consumption when system pressure is low to achieve demand side response. Through these measures, the system can effectively smooth the load fluctuation and improve the stability of the grid operation.

[0240] In this embodiment, the hierarchical intelligent scheduling strategy includes an hourly economic optimization control strategy:

[0241] S4.5.10: Identify the third frequency component in the hierarchical intelligent scheduling strategy, and use an improved back propagation learning rule to construct a multi-objective optimization function, including: minimum energy cost, minimum carbon emissions, and maximum reliability of electricity, gas, and heat energy;

[0242] The identification process of the third frequency component is specifically tailored to the needs of hourly-level economic optimization control, which mainly reflects the long-term operation trends and economic optimization opportunities of the system. Unlike second-level and minute-level control, hourly-level control focuses more on the overall economic performance and long-term sustainability of the system, requiring global optimization in a larger time and space range. The third frequency component usually contains information such as daily cycle changes of load, long-term fluctuations of renewable energy, changing trends of fuel prices, and fluctuations of power market prices.

[0243] The improved backpropagation learning rule plays a key role in the construction of multi-objective optimization functions. Traditional backpropagation algorithms mainly optimize a single objective, while improved versions can handle multiple conflicting optimization objectives simultaneously. This improvement includes multi-gradient calculation, adaptive adjustment of objective weights, and search for Pareto optimal solutions. The improved learning rule can better handle the nonlinear relationships and constraints between objective functions, avoid falling into local optimal solutions, and improve the effect of global optimization.

[0244] The objective of minimizing the operating cost of the power system covers multiple aspects such as fuel cost, equipment maintenance cost, start-stop cost, and electricity purchase cost. Fuel cost usually accounts for the largest proportion of total operating cost, including the consumption of coal, natural gas, nuclear fuel, etc. Equipment maintenance cost is related to the operation time and load level of the equipment, and frequent start-stop and high load operation will increase the maintenance demand. Start-stop cost is a cost item specific to traditional power generation equipment such as thermal power units, reflecting the additional consumption during the start-up and shutdown process. Electricity purchase cost comes from power market transactions or power exchange with other power grids.

[0245] Optimization of natural gas system operating costs needs to consider gas source procurement costs, pipeline transportation costs, compressor operating costs, and storage facility operating costs. Gas source procurement costs are greatly affected by fluctuations in the natural gas market price, and the system needs to have the ability to predict prices and optimize procurement strategies. Pipeline transportation costs are related to transportation distance, pipeline pressure, and flow size. Compressor operating costs mainly come from power consumption, which needs to be considered in coordination with the power system. The operation of the storage facility involves the selection of gas injection and gas extraction timing, and needs to balance storage costs and gas supply flexibility.

[0246] Optimization of thermal system operating costs includes heat source fuel cost, heat network circulating pump operating cost, heat exchange station operating cost, and heat loss cost. Heat source fuel cost varies according to the type of heat source, including fuel costs of coal, gas, biomass, etc. The operating cost of the heat network circulating pump is mainly the power consumption, which needs to balance the heating quality and energy efficiency. The operating cost of the heat exchange station includes equipment maintenance and control system operating costs. Heat loss cost reflects the energy loss of the heat network during transmission, which can be reduced by optimizing operating parameters and improving insulation measures.

[0247] The carbon emission minimization objective reflects the system's emphasis on environmental protection and commitment to sustainable development. Different energy types have different carbon emission factors, with coal power having the highest carbon emission intensity, followed by natural gas power, and renewable energy having nearly zero carbon emission. The system can minimize carbon emissions while meeting energy demand by optimizing energy structure and operation mode. Carbon emission optimization also needs to consider indirect emissions during energy conversion, such as the environmental impact of power production on conversion processes such as electric heating and hydrogen production.

[0248] The system reliability maximization objective ensures the safety, stability, and service quality of energy supply. Reliability indicators include power supply reliability, gas supply reliability, and heat supply reliability, reflecting the system's ability to meet user demand. Reliability optimization needs to consider device redundancy configuration, backup capacity arrangement, and fault recovery capability. High reliability usually requires additional investment and operating costs, so a balance between reliability and economy needs to be found.

[0249] Specifically, the system identifies the low-frequency component (<0.1 Hz) in the scheduling strategy, mainly for economic optimization control. The system uses an improved back propagation (BP) learning rule to construct a multi-objective optimization function. The optimization objectives include: energy operation cost minimization min∑(CePe + CgFg + ChQh), where Ce, Cg, and Ch are the unit energy prices of electricity, gas, and heat, respectively, and Pe, Fg, and Qh are the electric power, gas flow, and heat power, respectively; carbon emission minimization min∑(αePe + αgFg + αhQh), where αe, αg, and αh are the carbon emission factors of electricity, gas, and heat, respectively; system reliability maximization max∑Ri, where Ri is the reliability index of each part of the system.

[0250] S4.5.11: Adopt a multi-objective gradient descent algorithm to optimize the weighted combination of the gradients of the multi-objective optimization function, and adjust the weight coefficients adaptively according to the operating state of the integrated energy system through a dynamic weight adjustment strategy to obtain the optimal scheduling parameters;

[0251] The application of the multi-objective gradient descent algorithm is the core technical means to solve the complex optimization problem of the integrated energy system. Unlike traditional single-objective optimization, multi-objective optimization needs to consider multiple possible conflicting objective functions simultaneously, seeking a solution that can achieve the best balance among the objectives. The algorithm calculates the gradient direction of each objective function, then combines these gradient vectors by weighting, forming a comprehensive search direction. The advantage of this method is that it can improve multiple objectives simultaneously, avoiding the need to pre-determine the priority of objectives in traditional methods.

[0252] The gradient calculation process involves the computation of partial derivatives of each objective function with respect to the decision variables. In a comprehensive energy system, the decision variables can include the output levels of various power generation devices, the charging and discharging power of energy storage devices, the operating modes of energy conversion devices, and hundreds of other variables. The gradient calculation of each objective function needs to consider the complex coupling relationships between variables and nonlinear constraint conditions. The system uses automatic differentiation techniques and numerical gradient estimation methods to ensure the accuracy and efficiency of gradient calculation.

[0253] The weighted combination optimization strategy is a key link in the multi-objective gradient descent algorithm. The gradient vectors of different objective functions have different dimensions and numerical ranges, and direct linear combination may result in the neglect of the gradients of some objectives. The system uses normalization processing and adaptive scaling techniques to ensure that the gradients of each objective can play an appropriate role in the combination. The determination of the weighting coefficients is not static, but is dynamically adjusted according to the current system state and optimization progress.

[0254] The dynamic weight adjustment strategy reflects the intelligent and adaptive characteristics of the system. The weight adjustment strategy takes into account multiple factors, including system operating state, external environmental conditions, user demand changes, and market price fluctuations. During periods of tight power supply, the system automatically increases the weight of the reliability objective to ensure power supply safety; during periods of greater environmental pressure, the weight of the carbon emission minimization objective is correspondingly increased; during periods of rising fuel prices, the weight of the economic objective is strengthened.

[0255] The adaptive adjustment mechanism of the weight coefficient uses a variety of intelligent algorithms, including fuzzy logic control, neural network prediction, and reinforcement learning methods. Fuzzy logic control can handle the uncertainty and fuzziness in weight adjustment, and adjust the weights according to expert experience and operating rules. The neural network prediction method learns the relationship between weights and system state in historical data to predict the optimal weight configuration under the current conditions. The reinforcement learning method learns through interaction with the environment to gradually improve the weight adjustment strategy.

[0256] Comprehensive assessment of system operating state is the basis for weight adjustment. The assessment includes load level and its trend, renewable energy output, device health status, energy storage system state of charge, network congestion, and other dimensions. The system establishes a comprehensive state assessment index system that can quantitatively describe the overall operating condition of the system. Based on these state information, the weight adjustment strategy can make scientific and reasonable adjustment decisions.

[0257] The obtaining of optimal scheduling parameters is the ultimate goal of the multi-objective optimization process. These parameters not only need to meet the current optimization objectives, but also need to have the ability to adapt to future uncertainties. The determination of optimal parameters uses advanced methods such as robust optimization and stochastic optimization, considering various uncertainties such as load forecasting errors, renewable energy output uncertainties, and equipment failure risks. The final scheduling parameters obtained have good robustness and adaptability, and can maintain good performance under various operating conditions.

[0258] The multi-objective gradient descent algorithm is used to optimize the weighted combination of the gradients of the multi-objective optimization function, which is an advanced technical method for solving complex scheduling problems in integrated energy systems. The multi-objective gradient descent algorithm is an extension of the traditional gradient descent method in the context of multi-objective optimization, which considers multiple possible conflicting optimization objectives simultaneously. In integrated energy systems, these objectives usually include minimizing economic cost, maximizing energy utilization efficiency, minimizing environmental impact, and maximizing system reliability. Traditional single-objective optimization methods are difficult to balance these mutually restrictive objectives, while the multi-objective gradient descent algorithm finds a solution that balances all objectives by considering the gradient direction of all objective functions simultaneously in each iteration. Specifically, the algorithm first calculates the gradient of each individual objective function with respect to the scheduling parameters, which indicates how to adjust the parameters to improve the specific objective. Then, the algorithm combines these gradient vectors by weighting to form a comprehensive gradient direction. The key to weighted combination is how to determine the weight coefficients of each objective, which directly affects the final solution to which objectives it is biased. Traditional methods usually use fixed weights, but this approach is difficult to adapt to the dynamic characteristics of integrated energy systems. Therefore, the system introduces a dynamic weight adjustment strategy to adaptively adjust the weight coefficients according to the system's operating state. For example, when the system detects that a power peak is approaching, it may increase the weight of the power supply reliability objective; when environmental monitoring shows that air quality is declining, it may increase the weight of the emission reduction objective; when energy prices fluctuate significantly, it may adjust the weight of the economic objective. This dynamic adjustment ensures that the optimization process can respond to real-time demands of the system and changes in the external environment.

[0259] The dynamic weight adjustment strategy is a key mechanism for intelligent scheduling, as it adapts the weight coefficients based on the operating state of the integrated energy system. This strategy is based on a series of pre-set rules and intelligent algorithms, which can adjust the weight coefficients of each objective in multi-objective optimization in real-time, according to the system's operating state, external environmental conditions, and user preferences. This strategy considers multiple factors for weight adjustment. First, the system load state: during high load periods, the system may face supply pressure, and the weight of the reliability objective will automatically increase; during low load periods, the economy or environmental protection objective may have a higher weight. Second, renewable energy output: when renewable energy output is abundant, the system will increase the weight of the environmental protection objective to encourage more green energy consumption; when renewable energy output is unstable, the weight of the system stability objective will increase. Energy price fluctuations are also an important consideration: rising fuel prices may lead to an increase in the weight of the economic objective, encouraging the system to seek more cost-effective operation. In addition, external factors such as weather conditions, seasonal changes, and special events (such as major conferences and sports events) will also trigger weight adjustment. The system uses a variety of technologies to achieve this dynamic adjustment, including rule-based expert systems, fuzzy logic controllers, and reinforcement learning algorithms. The rule-based method uses pre-defined "if-then" rules to directly set the weight based on the system state; the fuzzy logic method can handle uncertainty and partial truth, more closely resembling the decision-making mode of human experts; reinforcement learning learns the optimal weight adjustment strategy through continuous interaction with the environment. The system also designs a stabilizing mechanism to prevent excessive weight fluctuations, using a sliding window average and gradual adjustment strategy to ensure that weight changes are smooth rather than abrupt.

[0260] The optimal scheduling parameters, which are the final goal of the multi-objective gradient descent and dynamic weight adjustment, directly guide the actual operation of the integrated energy system. The optimal scheduling parameters are a set of values that determine the operating state, output level, and control strategy of each device in the system at different time periods. For power generation devices, the parameters include start-stop state, output level, and ramp rate limit; for energy storage devices, they include charging and discharging power, time, and target energy storage level; for electric-thermal-gas conversion devices, they include conversion ratio and operating mode; for demand-side resources, they include load reduction amount and response time. These parameters are not determined in isolation, but as a whole, ensuring the coordinated operation of all parts of the system. During the parameter generation process, the system not only considers the current state, but also predicts the evolution of the system in the future period, achieving "forward-looking" scheduling. For example, the system may foresee a power consumption peak in a few hours and arrange energy storage charging or adjust the state of thermal storage devices in advance. To ensure the feasibility of the scheduling parameters, the system also considers various constraints, such as device physical limitations, network transmission capacity, environmental regulation requirements, etc. When multiple objectives cannot be simultaneously optimized, the system seeks a Pareto optimal solution, i.e., a solution that cannot improve any one objective without compromising at least one objective. The final generated scheduling parameters also undergo reliability and robustness tests, evaluating the adaptability of the parameters through simulation of different disturbance scenarios (such as load sudden changes, device failures, and renewable energy fluctuations). The system also provides parameter interpretation functions to explain the decision-making logic and expected effects behind each set of parameters to operating personnel, enhancing system transparency and credibility. These optimal scheduling parameters are delivered to each device control system through standard communication protocols, completing the closed loop from intelligent decision-making to actual control.

[0261] Specifically, the system uses a multi-objective gradient descent algorithm to solve the optimization function. The gradient update formula is For multi-objective gradient combination, λ i is the weight coefficient, represents the scheduling parameters (decision variables) at the kth iteration, including the output level of various power generation devices, the charging and discharging power of energy storage devices, the operating mode of energy conversion devices, etc. represents the scheduling parameters at the k+1th iteration (updated), represents the optimization objective function at the parameter , represents the gradient of the corresponding objective function , indicating how the parameters should be adjusted to improve the objective function, , and The first, second, and third optimization objective functions are respectively. The system adopts a dynamic weight adjustment strategy to adjust the weight coefficients according to the current operating state, for example, increasing the reliability weight during the peak load period, increasing the carbon emission weight when environmental protection requirements are high, and increasing the cost weight when economic pressure is great. Through this adaptive optimization, the system can obtain the optimal scheduling parameters.

[0262] S4.5.12: Constraint verification is performed on the optimal scheduling parameters, including device output limits, network transmission constraints, energy balance constraints, and a feasible scheduling scheme is generated based on the verification results;

[0263] Constraint verification is a key step to ensure the feasibility and safety of the scheduling scheme. In the complex operating environment of the integrated energy system, there are a large number of physical constraints, technical constraints, and safety constraints, and any scheduling scheme that violates the constraints may cause device damage, system instability, or supply interruption, etc. The constraint verification process needs to systematically check all relevant constraints to ensure that the generated scheduling parameters are technically feasible, economically reasonable, and safely reliable.

[0264] Device output limits are the most basic physical constraints, and each device has its designed maximum and minimum output range. For power generation devices, these limits include maximum power generation capacity, minimum stable output, ramp rate limit, continuous operation time limit, etc. For energy storage devices, constraints include maximum charge and discharge power, capacity limit, state of charge boundary, charge and discharge efficiency variation, etc. For energy conversion devices, constraints involve conversion efficiency curve, operating condition range, start-stop time limit, etc. The system needs to ensure that the scheduling instructions of all devices are within their technically feasible range.

[0265] Network transmission constraints reflect the physical limitations and safety requirements of the energy network. Transmission constraints of the power grid include line thermal stability limit, voltage stability limit, power angle stability constraint, short circuit capacity limit, etc. Constraints of the natural gas network involve pipeline flow limit, pressure constraint, compressor capacity limit, gas storage operation constraint, etc. Constraints of the heat network include pipeline transportation capacity, temperature and pressure limit, circulating pump capacity constraint, heat exchange capacity limit, etc. Verification of these constraints needs to consider the topology and operating state of the network.

[0266] Energy balance constraints are the basic requirements for system operation, ensuring that supply can meet demand at any time. Power balance requires that the sum of power generation and energy storage discharge equals the sum of load demand, energy storage charging, and network loss. Natural gas balance requires that the sum of gas supply and gas storage release equals the sum of gas demand, gas storage injection, and pipeline network loss. Heat balance requires that the heat source supply equals the heat load demand plus the net absorption of thermal storage devices. Verification of the balance constraints needs to consider the supply-demand matching of all energy forms.

[0267] The choice of constraint handling method has a significant impact on the optimization results and computational efficiency. The interior point method is an effective method for handling inequality constraints. By adding a logarithmic barrier term to the objective function, the constrained optimization problem is transformed into an unconstrained optimization problem. This method has the advantage of ensuring that the constraint conditions are always met during the iteration process, avoiding the generation of infeasible solutions. The Lagrange multiplier method is suitable for handling equality constraints. By introducing Lagrange multipliers, the constraint conditions are transformed into necessary conditions for the optimization problem.

[0268] The generation of a feasible scheduling scheme is the ultimate goal of the constraint verification process. When the initial optimization results violate certain constraints, the system needs to adjust and modify the scheduling parameters. The adjustment process uses an iterative optimization method to gradually modify the parameters that violate the constraints while maintaining the optimization objective as good as possible. This adjustment is not a simple parameter truncation, but a coordinated adjustment that takes into account the relationships between parameters. The final generated feasible scheme not only satisfies all the constraints, but also achieves a relatively optimal performance level within the feasible region.

[0269] Specifically, the system verifies the scheduling parameters obtained by optimization against the constraint conditions to ensure the feasibility of the scheme. The constraint conditions include: device output limit P min ≤ P ≤ P max , network transmission constraint |P ij | ≤ P ij,max (power grid)or |F ij | ≤F ij,max (gas network), energy balance constraint ∑P in = ∑P out + P loss , etc. The system handles these constraints through the interior point method or the Lagrange multiplier method to generate a feasible scheduling scheme that satisfies all the constraints.

[0270] S4.5.13: For the feasible scheduling scheme, develop hour-level scheduling plans for different devices, including generator start-stop timing, energy storage charging-discharging strategies, and combined heat and power operation modes, to obtain an economically optimized scheduling strategy.

[0271] The development of hour-level scheduling plans is a key process that converts abstract optimization results into specific operational instructions. This process needs to decompose the overall parameters in the feasible scheduling scheme into detailed operational plans for specific devices, ensuring that each device has clear operational instructions and time arrangements. The development of scheduling plans not only considers the characteristics of individual devices, but also coordinates the relationships and timing between different devices.

[0272] The arrangement of generator set start-stop timing is an important part of power system dispatching, directly affecting the economy and safety of the system. The start-stop timing needs to consider the minimum running time, minimum downtime, start-up cost, climbing ability and other technical constraints of the unit. For units that need to start, the system will arrange the start time according to the expected load demand and start-up time to ensure that the unit can be put into operation in time. For units that need to stop, the system will choose a period with low load and other units that can be replaced for shutdown operation.

[0273] The development of generator set output curve needs to consider load variation, economic requirements and technical constraints. The output curve is not a simple sequence of power values, but a continuous adjustment plan considering the dynamic characteristics of the unit. During the curve development process, it is necessary to ensure that the output change between adjacent periods does not exceed the climbing ability limit of the unit, while considering the efficiency change and environmental impact of the unit at different output levels.

[0274] The development of energy storage system charging and discharging strategy fully utilizes the bidirectional regulation ability and fast response characteristics of energy storage devices. The charging period is usually arranged in periods of low load or high renewable energy output, to store excess energy for future use. The discharge period is arranged in periods of high load or high generation cost, to release stored energy to reduce the use of expensive power sources. Strategy development also needs to consider the cycle life and efficiency loss of energy storage devices to avoid damage to the devices caused by excessive charging and discharging.

[0275] Energy storage power curve and capacity management plan ensure the effective operation of the energy storage system throughout the dispatching period. The power curve specifies the charging and discharging power level of each period, which needs to be optimized within the technical limits of the energy storage device. The capacity management plan focuses on the state of charge trajectory of the energy storage device to ensure sufficient available capacity at critical moments. The planning process also needs to reserve a certain amount of capacity for emergency situations.

[0276] The determination of combined heat and power unit operation mode needs to coordinate the coupling relationship between power output and heat output. Under different operation modes, the unit's electric-thermal ratio, total efficiency and economy will change. The system needs to choose the most suitable operation mode and output distribution according to the demand for electricity and heat. When the demand for electricity is high, the electrically determined heat operation mode may be used; when the demand for heat is high, the heat-determined electric mode may be used; when supply and demand are balanced, the highest efficiency operation point is selected.

[0277] The optimization of cogeneration power distribution ratio is a multi-objective decision-making process. The distribution process needs to consider multiple factors such as electricity market prices, heat supply costs, unit efficiency curves, environmental emissions, etc. The optimization objectives usually include minimizing total operating costs, maximizing energy utilization efficiency, minimizing environmental impact, etc. The determination of the distribution ratio not only affects the operation effect of the current period, but also affects the adjustment capacity and operation flexibility of the subsequent period.

[0278] The implementation of economic optimization scheduling strategy is the final embodiment of the whole hour-level control. The strategy implementation process includes scheduling instruction issuing, execution monitoring, deviation correction, etc. The instruction issuing adopts standardized communication protocol to ensure that all devices can accurately receive and understand the scheduling instruction. The execution monitoring discovers deviations and problems in the execution process through real-time data acquisition and analysis. The deviation correction mechanism can quickly respond when detecting execution deviation, and maintain the scheduling effect through backup scheme or parameter adjustment.

[0279] The dynamic adjustment mechanism of the scheduling strategy ensures that the strategy can adapt to various changes in actual operation. When the prediction error is large or a sudden event occurs, the system can quickly recalculate and adjust the scheduling plan. The adjustment process is not a simple parameter modification, but a re-optimization based on the current state, ensuring that the adjusted scheme still meets all constraint conditions and maintains good economic performance. This dynamic adjustment capability is an important feature of intelligent scheduling of modern integrated energy systems.

[0280] Specifically, the system formulates specific hour-level scheduling plans for different devices based on feasible scheduling schemes. This includes the start-stop timing and output curve of generator units, the charging-discharging time and power curve of energy storage systems, the operation mode and output distribution ratio of cogeneration units, etc. These scheduling plans are issued to the control systems of each device through the communication network to implement economic optimization scheduling strategy and achieve efficient operation of the system.

[0281] In this embodiment, the execution results of the hierarchical intelligent scheduling strategy are monitored in real time, and a fault-tolerant control mechanism is constructed using the belief propagation ordered statistical decoding BP-OSD algorithm. When the operation deviation of the integrated energy system exceeds the preset threshold, error detection and correction are performed, including:

[0282] S5.1: Real-time data acquisition of the execution results of the hierarchical intelligent scheduling strategy, and encoding different device states and energy flow data into check code words with error correction capability according to BCH code or LDPC code rules to obtain encoded data of the operation state of the integrated energy system;

[0283] Specifically, the system collects key operating parameters such as electric power Pe, gas flow Fg, thermal power Qh, voltage V, frequency f in real time with a period of 100 ms, eliminates measurement noise through Kalman filter, and obtains high-precision state estimation values. These data are converted into code words with error correction capability according to the encoding rules of BCH code or LDPC code. The encoding process is c = mG, where m is the information sequence, G is the generator matrix, and c is the encoded code word. Through this encoding mode, the system can detect and correct possible errors in subsequent processing.

[0284] S5.2: For the encoded data, a belief propagation algorithm is used for error detection, and a plurality of rounds of iteration calculation are performed through factor graph representation and message passing from variable nodes to check nodes to obtain error position probability distribution;

[0285] The use of a belief propagation algorithm for error detection on encoded data is an important technical means to ensure the reliability of data in a comprehensive energy system. The belief propagation algorithm is an iterative algorithm for probability reasoning, first proposed by Pearl in 1982, and is a basic algorithm in the field of artificial intelligence and information theory. In the field of communication and signal processing, belief propagation is widely used in decoding error correction codes. The core idea of the belief propagation algorithm is to pass and update information (or "belief", "confidence") on a probabilistic graph model, and finally obtain a global optimal or approximately optimal solution through local calculation. In a comprehensive energy system, data may be disturbed and produce errors during transmission, storage and processing. If these errors are not discovered and corrected in time, they may lead to incorrect control decisions and affect system safety and efficiency. The belief propagation algorithm realizes efficient error detection by representing the relationship between encoded data through a factor graph. The factor graph is a bipartite graph containing two types of nodes: variable nodes and factor nodes. In error detection applications, variable nodes represent each bit in the encoded data, and factor nodes represent the constraint relationships (such as check equations) between these bits. Information is passed between variable nodes and factor nodes in the form of probability, reflecting each node's "belief" about the state of other nodes. Variable nodes send messages to connected factor nodes expressing their probability distribution of state; factor nodes receive messages from all connected variable nodes, and after comprehensive calculation, they send updated messages back to each variable node. This message passing process usually requires multiple rounds of iteration until the beliefs of all nodes converge or reach a preset number of iterations. In each round of iteration, each node only updates its belief based on the messages from its neighbor nodes without needing to understand the global state of the entire graph, which makes the algorithm efficient even when dealing with large-scale problems. Through iterative calculation, the belief propagation algorithm finally outputs the probability estimate of each bit being wrong, forming an error position probability distribution, which provides an important basis for the subsequent error correction process.

[0286] The core mechanism of belief propagation algorithm is the multi-round iterative computation through factor graph representation and message passing between variable nodes and check nodes. Factor graph is a special probabilistic graphical model used to visually represent the dependencies between variables in complex systems. In error detection of multi-energy systems, factor graph consists of two types of nodes: variable nodes and check nodes (also known as factor nodes). Variable nodes represent each bit or symbol in the data, and the possible values are usually 0 or 1; check nodes represent the constraints imposed by the coding rules, such as parity check, etc. Edges connect variable nodes and check nodes, representing that a variable participates in a certain constraint condition. Message passing is the core operation of belief propagation algorithm, including two directions of transmission: from variable nodes to check nodes, and from check nodes to variable nodes. The message passing from variable nodes to check nodes is essentially the variable nodes passing their own probability information (excluding information from the target check node) to the check nodes. This message represents "the probability of the variable node being in various possible states based on all other information sources". The message passing from check nodes to variable nodes is the check nodes calculating the probability of the target variable node being in various states based on all other variable nodes' messages received and the constraint condition itself. This two-way message passing forms an iterative process, and with the increase of iteration times, the estimation of data state by each node gradually becomes accurate. In practical systems, iteration usually continues until a certain stopping condition is met, such as belief change less than a threshold, reaching the maximum number of iterations, or the decoding result satisfying all check equations. The key advantage of multi-round iteration is that the propagation of information is not limited to directly connected nodes, but can be transmitted through intermediate nodes to the entire network, achieving global information integration. This iterative computation finally outputs the probability estimate of each bit error, i.e. the error location probability distribution, providing a basis for subsequent error correction.

[0287] Error location probability distribution is a key result output by belief propagation algorithm, which reflects the possibility of each position in the data appearing error. This distribution is usually expressed in the form of probability values, with a value range from 0 to 1, where a higher probability value indicates that the position is more likely to appear error. In the data processing of multi-energy systems, error location probability distribution provides an important basis for decision-making for subsequent error correction process, enabling the system to focus on correcting the most likely error position rather than blindly checking all positions. Order statistics decoding algorithm is an efficient method for error correction based on error location probability distribution. The algorithm first sorts all positions according to error probability from high to low, and then strategically generates multiple candidate correction schemes. The core idea of order statistics decoding algorithm is "correct the most likely error position first", which is consistent with human intuition: when we find a spelling error in a text, we usually prefer to check those words that look most suspicious. When generating candidate correction schemes, the algorithm uses various strategies, such as flipping different numbers of high-probability error bits one by one, combining domain knowledge for targeted correction, and considering the correlation of error patterns, etc. These strategies ensure that the generated candidate schemes cover various possible error cases and avoid the computational explosion caused by enumerating all possible combinations. For each generated candidate correction scheme, the algorithm needs to evaluate its probability of being the correct decoding. This is usually achieved by calculating the Euclidean distance and likelihood probability of the candidate scheme and the received sequence. The Euclidean distance is an indicator that measures the "straight-line distance" between two vectors, which measures the closeness of the candidate scheme and the received data in the signal space in this context. The likelihood probability measures the probability of the candidate scheme being the original transmitted data under the condition of receiving the current sequence under certain channel conditions.

[0288] Specifically, the system uses the belief propagation (BP) algorithm for error detection on the received encoded data r. First, a factor graph G = (V, C, E) is constructed, where V is the set of variable nodes, C is the set of check nodes, and E is the set of edges. The message update formula from variable node to check node is q ij (x i ) = α ij ·p i (x i )·∏k∈N(i)\j rk j (x i ), and the message update from check node to variable node is r ji (x i ) = ∑{xk:k∈N(j)\i} ∏k∈N(j)\i qk j (xk). The system obtains the posterior probability P(x i |r) of each bit through 5-10 rounds of iterative calculation (until the posterior probability changes less than 10 -4 between adjacent two rounds), thereby determining the most likely error position.

[0289] S5.3: Based on the error location probability distribution, generate multiple candidate error correction schemes using an ordered statistic decoding algorithm, calculate the Euclidean distance and likelihood probability of each candidate error correction scheme with the received sequence, select the candidate error correction scheme with the maximum likelihood probability for error correction, and obtain the integrated energy system operation deviation correction instruction.

[0290] The integrated energy system operation deviation correction instruction is the final output of the entire error detection and correction process, and is also the key command for actual system operation. These instructions provide accurate correction measures for the detected deviations in system operation, ensuring that the system returns to the optimal operating state. In a multi-energy system, operational deviations can be caused by various factors: equipment failure, load prediction error, renewable energy output fluctuation, communication delay or packet loss, etc. If these deviations are not corrected in time, it may lead to energy supply and demand imbalance, system efficiency decline or safety boundary approach. Deviation correction instructions usually contain multiple aspects of information: the target device or subsystem for correction, the direction and amplitude of correction, the execution time window, priority, etc. For example, to correct the power supply shortage deviation, the instruction may include increasing the output of a specific generator, starting a backup unit, invoking demand response resources to reduce load, etc.; to correct the low heating temperature problem of the heat supply network, the instruction may include adjusting the water supply temperature of the heat source device, increasing the circulating water flow, optimizing the operating parameters of the heat exchange station, etc. These instructions are not generated in isolation, but are optimized as a whole, taking into account various constraint conditions and objective functions. The generation process of deviation correction instructions usually uses the Model Predictive Control (MPC) method, which not only considers the current deviation, but also predicts the evolution of the system in the future, achieving forward-looking regulation. These instructions are issued to each execution device through standard communication protocols, completing the closed loop from information processing to physical control. To ensure the reliable execution of the instructions, the system also sets up a multi-level safety check mechanism, such as reasonableness verification, physical constraint verification, execution feedback confirmation, etc. In some critical scenarios, the instruction execution also adopts the "human-in-the-loop" method, i.e. important instructions need to be confirmed by humans before execution, adding an extra layer of safety protection. Through these carefully designed deviation correction instructions, the multi-energy system can respond to various operational disturbances in a timely manner, maintaining an efficient, reliable and economic operating state, and maximizing the overall advantages of the integrated energy system.

[0291] Specifically, based on the error location probability distribution obtained by the BP algorithm, the OSD algorithm is used to generate candidate error correction schemes. First, the reliability of the received sequence is sorted, and the most unreliable η positions (usually η = 16 ~ 64) are selected to generate error patterns, resulting in 2 η candidate codewords. For each candidate codeword, calculate its Euclidean distance dE = ∑(ri - c i )²and likelihood probability L(c|r) = exp(-dE² / 2σ²), the scheme with the maximum likelihood probability is selected as the final error correction result, where r i represents the value of the i-th position of the received sequence, i.e. the actual received (possibly containing errors) encoded data, c i represents the value of the i-th position of the candidate codeword, i.e. the possible correct codeword generated by the error correction algorithm. The system generates specific control instructions according to the error correction result, including generator set output adjustment ΔPG, energy storage charge and discharge power adjustment ΔPB, controllable load switching ΔPL, etc., and issues these instructions to each execution device through the communication network to correct the operation deviation of the integrated energy system.

[0292] Generating multiple candidate error correction schemes using ordered statistical decoding algorithm is an efficient error correction strategy. Ordered statistical decoding (OSD) was first proposed by Fossorier and Lin in 1995, and is a soft decision decoding algorithm, especially suitable for short to medium length linear codes. Compared with traditional hard decision decoding method, OSD can utilize soft information (i.e. reliability index) to significantly improve decoding performance. In the multi-energy system, OSD algorithm first sorts all bits according to the error position probability distribution to identify the most unreliable k positions (k is usually a preset parameter). Then, the algorithm generates 2 k candidate error correction schemes, each corresponding to a possible combination of the k positions. For example, if k=3, there are 8 possible combinations. In practical applications, in order to reduce computational complexity, improved OSD algorithm is usually used, such as considering only the case of flipping no more than t bits.

[0293] Computing the Euclidean distance and likelihood probability between each candidate correction scheme and the received sequence is a critical step in evaluating the quality of the candidate schemes. The Euclidean distance originates from Euclidean geometry and is the straight-line distance between two points in a multi-dimensional space. In the field of signal processing, the Euclidean distance is often used to measure the similarity between two signals or sequences. For binary data, the Euclidean distance can be simplified to the Hamming distance, which is the number of different bits at corresponding positions in two sequences. When calculating the Euclidean distance, the system treats the candidate correction scheme and the received sequence as two points in a multi-dimensional space and then calculates the straight-line distance between them. The smaller the distance, the more similar the two sequences are, and the more likely the candidate scheme is close to the original transmitted data. The likelihood probability is a more complex but also more accurate evaluation indicator. The likelihood probability refers to the conditional probability of receiving the current sequence if the candidate correction scheme is transmitted under given channel conditions. This probability depends on the characteristics of the communication channel, such as noise distribution, interference pattern, etc. In a Gaussian noise channel, the likelihood probability has an exponential relationship with the Euclidean distance, and the smaller the distance, the greater the likelihood probability. However, in more complex channel models, such as Rayleigh fading channels or channels with burst error characteristics, the calculation of the likelihood probability needs to consider more factors. In actual systems, the calculation of the likelihood probability is usually based on statistical modeling of the channel and analysis of historical data. The system will collect and analyze the historical patterns of communication errors to establish a channel model suitable for specific application scenarios. Based on this model, the system can calculate the likelihood probability for each candidate correction scheme and select the scheme with the maximum likelihood probability as the final decoding result. This statistical-based method enables the system to adapt to different communication environments and error patterns, improving the correction ability and robustness.

[0294] It should be noted that selecting the candidate error correction scheme with the maximum likelihood probability for error correction is the last step of the decoding process and is a key link to determine the error correction performance. Maximum likelihood decoding (MLD) is a basic concept in information theory, proposed by Shannon in the 1940s. The core idea is: among all possible code words, select the one with the highest probability of receiving sequence as the decoding result. This principle can minimize the decoding error rate in theory and is the optimal decoding strategy. In practical applications, since the number of candidate error correction schemes can be very large, it is usually not feasible to exhaust all possibilities. This is why the previous step uses algorithms such as ordered statistical decoding to generate a limited number of high-quality candidate schemes, rather than trying all possible combinations. In the final selection stage, the system compares the likelihood probabilities of all candidate schemes and selects the one with the highest probability as the decoding result. If multiple schemes have similar likelihood probabilities, the system may introduce additional decision criteria, such as considering consistency with historical data, prior knowledge in the application field, etc. Once the final error correction scheme is determined, the system will correct the errors in the original data accordingly to obtain more accurate information. In a multi-energy system, this corrected data is used to generate comprehensive energy system operation deviation correction instructions. These instructions directly affect the system's operation state adjustment, such as correcting the output level of power generation equipment, adjusting the charge and discharge strategy of energy storage equipment, optimizing the working mode of energy conversion equipment, etc. Accurate error correction is crucial to ensure the reliability of these instructions, as incorrect instructions can lead to resource waste, system efficiency decline, and even safety accidents. By combining belief propagation algorithm and ordered statistical decoding, the system can effectively detect and correct errors in the communication process, ensuring the accuracy and reliability of control instructions, and providing protection for the stable and efficient operation of the multi-energy system.

[0295] The application also provides a comprehensive energy system intelligent scheduling and control device, comprising:

[0296] A multi-scale modeling module is configured to acquire electric, gas, and heat resource operation data of a comprehensive energy system, and perform multi-scale hierarchical modeling on the operation data according to second-level, minute-level, and hour-level time scales and local cluster and regional cluster spatial scales.

[0297] A feature representation module is configured to perform feature extraction and dimension reduction representation on an electric resource state space model, a gas resource flow continuity model, a heat resource heat balance model, and a multi-energy coupling characteristic constraint model by using a deep autoencoder network.

[0298] A collaborative training module is configured to perform multi-group cognitive model collaborative training on resource representation data in a first dimension feature space by using a framework combining split learning and hierarchical federated learning.

[0299] a strategy generation module configured to construct a neural architecture search network with mixed bionic learning rules according to a global intelligent model with multiple groups of device collaborative cognitive capabilities, and generate a hierarchical intelligent scheduling strategy;

[0300] a fault-tolerant control module configured to monitor an execution result of the hierarchical intelligent scheduling strategy in real time, and construct a fault-tolerant control mechanism using a belief propagation ordered statistical decoding (BP-OSD) algorithm.

[0301] The functions and roles of the above modules correspond to the foregoing method embodiments, and will not be described here again.

[0302] According to the above, the method and device for intelligent scheduling and control of a comprehensive energy system provided by the present application realize multi-time scale collaboration, collaborative learning under multi-device group privacy protection, and fault-tolerant control with high reliability of the comprehensive energy system, and improve the overall operation efficiency and reliability of the system.

[0303] The above only describes preferred embodiments of the present application. It should be noted that, for those skilled in the art, without departing from the principles of the present application, several improvements and refinements can be made, which should also be considered as the protection scope of the present application.

[0304] The computer program is stored on a computer readable storage medium and is run by a processor to execute the steps described in the above method embodiments. The storage medium can be a volatile or non-volatile computer readable storage medium.

[0305] In addition, the present disclosure also provides a computer program product, which stores a computer program. When the computer program is run by a processor, it executes the steps provided by any of the above embodiments of the present disclosure. For details, please refer to the above method embodiments, which will not be described here again.

[0306] The computer program product can be implemented by hardware, software or a combination thereof. In one optional embodiment, the computer program product is specifically embodied as a computer storage medium, which can be a volatile or non-volatile computer readable storage medium. In another optional embodiment, the computer program product is specifically embodied as a software product, such as a software development kit (SDK) and the like.

[0307] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the device and the apparatus described above can refer to the corresponding process in the foregoing method embodiment, and will not be repeated here. In several embodiments provided in the present disclosure, it should be understood that the disclosed device, apparatus and method can be implemented in other ways. The apparatus embodiments described above are only schematic, for example, the division of the units is only a logical function division, and there can be another division in actual implementation, for example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the units shown or discussed can be indirect coupling or communication connection through some communication interfaces, devices or units, and can be electrical, mechanical or other forms.

[0308] The units described as separate components can or can not be physically separate, and the components shown as units can or can not be physical units, i.e., they can be located in one place or distributed on multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the embodiment.

[0309] In addition, each functional unit in each embodiment of the present disclosure can be integrated into a processing unit, or each unit can exist physically, or two or more units can be integrated into one unit.

[0310] If the functions are realized in the form of software functional units and sold or used as independent products, they can be stored in a non-volatile computer readable storage medium executable by a processor. Based on this understanding, the technical solutions of the present disclosure essentially or the part of the prior art or the part of the technical solutions can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a plurality of instructions to make a computer device (which can be a personal computer, a server, or a network device, etc.) execute all or part of the steps of the method described in each embodiment of the present disclosure. The foregoing storage medium includes: U disk, mobile hard disk, read-only memory (Read-Only Memory, ROM), random access memory (Random Access Memory, RAM), magnetic disk or optical disk, and various program code storage media.

[0311] Finally, it should be noted that the above-described embodiments are merely specific embodiments of the present disclosure, used to illustrate the technical solutions of the present disclosure, and are not intended to limit the present disclosure. The protection scope of the present disclosure is not limited thereto. Although the present disclosure has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can make modifications or easy changes to the technical solutions described in the foregoing embodiments, or easily think of changes or equivalent replacements for some of the technical features; and these modifications, changes or replacements do not cause the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present disclosure, and should be covered within the protection scope of the present disclosure. Therefore, the protection scope of the present disclosure should be subject to the protection scope of the claims.

Claims

1. An integrated energy system intelligent scheduling and control method, characterized in that, The method comprises the following steps: obtaining the operation data of the electricity, gas and heat resources of the comprehensive energy system, and performing multi-scale hierarchical modeling on the operation data according to the time scales of seconds, minutes and hours and the spatial scales of local clusters and regional clusters to obtain an electric resource state space model, a gas resource flow continuity model, a thermal resource heat balance model and a multi-energy coupling characteristic constraint model; performing feature extraction and dimension reduction representation on the electric resource state space model, the gas resource flow continuity model, the thermal resource heat balance model and the multi-energy coupling characteristic constraint model by using a deep auto-encoder network to obtain resource representation data in a first dimension feature space; performing collaborative training of multiple cognitive models on the resource representation data in the first dimension feature space by using a split hierarchical federated learning framework, performing parameter aggregation and global optimization of the multiple cognitive models based on the hierarchical architecture to obtain a global intelligent model with collaborative cognitive ability of multiple devices; constructing a neural architecture search network with hybrid bionic learning rules in the global intelligent model, setting different scheduling targets of second-level frequency stability, minute-level load suppression and hour-level economic optimization, and performing optimal bionic learning rule combination search in the neural architecture search network to generate and execute a hierarchical intelligent scheduling strategy, which comprises: in the global intelligent model, constructing the neural architecture search framework by using a reinforcement learning controller based on a long short-term memory network, taking the current network architecture code, historical performance indicators and system operation conditions as a state space to obtain a neural network architecture search state representation; based on the neural network architecture search state representation, sampling candidate architectures by using the reinforcement learning controller, generating candidate neural network architectures through an action space of layer selection, activation function type, connection mode and learning rule combination to obtain a diversified candidate neural network architecture set; performing neural network architecture performance evaluation on the diversified candidate neural network architecture set on a historical running data validation set, calculating a reward signal through comprehensive indicators of scheduling accuracy, response speed and energy efficiency to obtain a neural network architecture performance evaluation result; according to the neural network architecture performance evaluation result, optimizing the controller policy parameters by using a policy gradient algorithm, guiding the architecture search direction through the different scheduling targets of second-level frequency stability, minute-level load suppression and hour-level economic optimization to obtain an optimal bionic learning rule combination; for the optimal bionic learning rule combination, routing different frequency components to the corresponding bionic learning layer through a time scale decomposer, performing multi-layer information fusion by using an attention mechanism to generate and execute the hierarchical intelligent scheduling strategy; monitoring the execution result of the hierarchical intelligent scheduling strategy in real time, constructing a fault-tolerant control mechanism by using a belief propagation ordered statistical decoding BP-OSD algorithm, and performing error detection and correction when the operation deviation of the comprehensive energy system exceeds a preset threshold.

2. The method of claim 1, wherein, The operation data of electricity, gas and heat resources of the integrated energy system is obtained, and the operation data is subjected to multi-scale hierarchical modeling according to second-level, minute-level and hour-level time scales and local cluster and regional cluster spatial scales, to obtain an electric resource state space model, a gas resource flow continuity model, a heat resource heat balance model and a multi-energy coupling characteristic constraint model, including: For the electric resource operation data, a differential equation of a state vector and a control input is used to establish a dynamic characteristic modeling of electric resource charging and discharging, to obtain a relationship model of an energy storage state vector and a control input; For the gas resource operation data, a continuity constraint of gas density and flow rate is used to model gas flow, to obtain a gas density and flow rate distribution model; According to the heat resource operation data, the energy storage state vector, the control input relationship model, the gas density and the flow rate distribution model, a heat transfer process is modeled by using a balance relationship of mass heat capacity and temperature change, to build a constraint relationship of electric power, gas power and heat power of a multi-energy coupling device, to obtain the electric resource state space model, the gas resource flow continuity model, the heat resource heat balance model and the multi-energy coupling characteristic constraint model.

3. The method of claim 1, wherein, For the electric resource state space model, the gas resource flow continuity model, the heat resource heat balance model and the multi-energy coupling characteristic constraint model, a deep auto-encoder network is used for feature extraction and dimension reduction representation, to obtain resource representation data in a first dimension feature space, including: The electric resource state space model, the gas resource flow continuity model, the heat resource heat balance model and the multi-energy coupling characteristic constraint model are subjected to data preprocessing and normalization, to obtain standardized multi-dimensional operation state data; Based on the standardized multi-dimensional operation state data, an encoder network is constructed for nonlinear mapping of second dimension state data, to obtain a compressed intermediate representation vector, the second dimension being higher than the first dimension; A decoder network is used to reconstruct and verify the compressed intermediate representation vector, and the encoder parameters are optimized by minimizing the reconstruction error, to obtain the resource representation data in the first dimension feature space.

4. The method of claim 1, wherein, The hierarchical architecture includes a first layer network deployed on a local client, a second layer network deployed on an edge server and a third layer network deployed on a cloud server, and the split hierarchical federated learning framework is obtained by fusing split learning and hierarchical federated learning. For the resource representation data in the first dimension feature space, the split hierarchical federated learning framework is used for collaborative training of multiple cognitive models, and the parameter aggregation and global optimization of the multiple cognitive models are performed based on the hierarchical architecture, to obtain a global intelligent model with multiple device collaborative cognitive ability, including: For the resource representation data in the first dimension feature space, preliminary feature learning is performed in the first layer network deployed on the local client, and data privacy is protected by differential privacy and homomorphic encryption technology, to obtain an encrypted basic feature vector. Based on the encrypted basic feature vector, deep feature learning is performed in the second layer network deployed on the edge server, and cross-device correlation features are learned by using an attention mechanism and a graph convolution network to obtain a regional collaborative representation result. In the third layer network deployed on the cloud server, the regional collaborative representation result is aggregated and optimized by using a federated average algorithm, and a cross-region coordination mode is learned by splitting learning to obtain a global intelligent model with multi-group device collaborative cognitive ability.

5. The method of claim 4, wherein, Based on the encrypted basic feature vector, deep feature learning is performed in the second layer network deployed on the edge server, and cross-device correlation features are learned by using an attention mechanism and a graph convolution network to obtain a regional collaborative representation result, including: In the second layer network deployed on the edge server, the encrypted basic feature vector is fused by a four to six layer deep network, the correlation weight between different devices is calculated by a multi-head attention mechanism, and a cross-device coupling feature is obtained. According to the cross-device coupling feature, a graph convolution network is used to learn the load transfer path and the fault propagation mode, and the time-dependent relationship of the different devices is identified by time series modeling to obtain a spatio-temporal correlation mode. The spatio-temporal correlation mode is parameter aggregated and locally optimized in the second layer network, and the parameters of the second layer network are updated according to the client data weight and the loss function gradient by using a weighted average method to obtain the regional collaborative representation result.

6. The method of claim 1, wherein, The hierarchical intelligent scheduling strategy includes a second frequency component: The first frequency component in the hierarchical intelligent scheduling strategy is identified and extracted, and the frequency deviation signal is converted into a pulse firing rate proportional to the deviation amplitude by a Poisson encoder to obtain a frequency deviation pulse sequence. Based on the frequency deviation pulse sequence, a leaky integrate-and-fire model of a spiking neural network is used for membrane potential dynamic modeling, and the membrane potential change is calculated by the time constant, resting potential and input resistance of the membrane potential to obtain a pulse response signal. The pulse response signal is threshold judged, and a control pulse is generated and the membrane potential is reset when the membrane potential exceeds the firing threshold to obtain a frequency regulation instruction. According to the frequency regulation instruction, millisecond-level power regulation is performed on the generator set and energy storage device to eliminate frequency deviation by primary frequency modulation control. After eliminating the frequency deviation, the frequency control state is monitored and fed back in real time, and the frequency stability error is controlled within the range of plus or minus 0.05 Hz.

7. The method of claim 1, wherein, The hierarchical intelligent scheduling strategy includes a second frequency component: The second frequency component in the hierarchical intelligent scheduling strategy is extracted, and the adaptive ability of load prediction is enhanced by using Hebb learning rule, and the load change mode is learned by simultaneously activating the connection strength of neurons to obtain an adaptive load prediction model. According to the adaptive load prediction model, a time window mechanism is introduced to learn the periodic pattern and mutation characteristics of the load, and the prediction network parameters are trained by historical load data to obtain a load prediction result. The competitive learning mechanism is used for feature selection on the load prediction result, and the key feature recognition capability is enhanced through lateral inhibition, so that the load prediction accuracy is improved to more than 95%, and a high-precision load prediction result is obtained. Based on the high-precision load prediction result, the load deviation and fluctuation trend are calculated, the power compensation and peak load shifting are performed by using the energy storage system and controllable load, and a load suppression control strategy is obtained.

8. The method of claim 1, wherein, The hierarchical intelligent scheduling strategy includes an hourly economic optimization control strategy: The third frequency component in the hierarchical intelligent scheduling strategy is identified, and a multi-objective optimization function is constructed by using an improved back propagation learning rule, including: minimization of electric, gas and heat energy operation cost, minimization of carbon emission and maximization of reliability; A multi-objective gradient descent algorithm is used to optimize the gradient of the multi-objective optimization function, and the weight coefficient is adjusted adaptively according to the operation state of the integrated energy system by using a dynamic weight adjustment strategy, so as to obtain optimal scheduling parameters; The optimal scheduling parameters are verified under the constraint conditions, including device output limit, network transmission constraint and energy balance constraint, and a feasible scheduling scheme is generated based on the verification result; For the feasible scheduling scheme, an hourly scheduling plan for different devices is formulated, including generator set start-stop timing, energy storage charging and discharging strategy, and cogeneration operation mode, to obtain an economic optimization scheduling strategy.

9. The method of claim 1, wherein, The execution result of the hierarchical intelligent scheduling strategy is monitored in real time, and a fault-tolerant control mechanism is constructed by using a belief propagation ordered statistical decoding BP-OSD algorithm, and error detection and correction are performed when the operation deviation of the integrated energy system exceeds a preset threshold, including: Real-time data acquisition is performed on the execution result of the hierarchical intelligent scheduling strategy, and different device states and energy flow data are encoded into check code words with error correction capability according to BCH code or LDPC code rules, to obtain the encoded data of the operation state of the integrated energy system; For the encoded data, an error detection is performed by using a belief propagation algorithm, and a plurality of candidate error correction schemes are generated by using an ordered statistical decoding algorithm based on the error position probability distribution, and the Euclidean distance and likelihood probability of each candidate error correction scheme and the received sequence are calculated, and the candidate error correction scheme with the maximum likelihood probability is selected for error correction, to obtain the integrated energy system operation deviation correction instruction. The acquisition module is used to acquire the electric, gas and heat resource operation data of the integrated energy system, and the operation data is modeled in multiple scales according to the second, minute, hourly time scales and local cluster, regional cluster spatial scales, to obtain an electric resource state space model, a gas resource flow continuity model, a heat resource heat balance model and a multi-energy coupling characteristic constraint model.

10. An integrated energy system intelligent dispatching and control device, characterized in that, ​ ​ The feature extraction and dimension reduction representation module is configured to perform feature extraction and dimension reduction representation on the electrical resource state space model, the gas resource flow continuity model, the thermal resource heat balance model, and the multi-energy coupling characteristic constraint model by using a deep auto-encoder network to obtain resource representation data in a first dimension feature space; The collaborative training module is configured to perform collaborative training on the resource representation data in the first dimension feature space by using a split hierarchical federated learning framework to obtain a global intelligent model with multi-device collaborative cognitive capability by performing parameter aggregation and global optimization of the multiple cognitive models based on the hierarchical architecture. The search module is configured to construct a neural architecture search network with a hybrid bionic learning rule in the global intelligent model, set different scheduling targets of second-level frequency stability, minute-level load suppression, and hour-level economic optimization, and perform optimal bionic learning rule combination search in the neural architecture search network to generate and execute a hierarchical intelligent scheduling strategy. In the global intelligent model, a reinforcement learning controller based on a long short-term memory network is used to construct the neural architecture search framework, and the current network architecture, historical performance indicators, and system operating conditions are used as a state space to obtain a neural network architecture search state representation. Based on the neural network architecture search state representation, the reinforcement learning controller is used to sample candidate architectures, generate candidate neural network architectures through an action space of layer selection, activation function type, connection mode, and learning rule combination, and obtain a diversified candidate neural network architecture set. The diversified candidate neural network architecture set is evaluated for neural network architecture performance on a historical running data validation set, a reward signal is calculated through comprehensive indicators of scheduling accuracy, response speed, and energy efficiency, and a neural network architecture performance evaluation result is obtained. According to the neural network architecture performance evaluation result, a policy gradient algorithm is used to optimize controller policy parameters, different scheduling targets of second-level frequency stability, minute-level load suppression, and hour-level economic optimization are used to guide the architecture search direction, and an optimal bionic learning rule combination is obtained. For the optimal bionic learning rule combination, a time scale decomposer is used to route different frequency components to the corresponding bionic learning layer, an attention mechanism is used for multi-layer information fusion, and the hierarchical intelligent scheduling strategy is generated and executed. The monitoring module is configured to monitor the execution result of the hierarchical intelligent scheduling strategy in real time, construct a fault-tolerant control mechanism by using a belief propagation ordered statistical decoding (BP-OSD) algorithm, and perform error detection and correction when the operating deviation of the integrated energy system exceeds a preset threshold.

Images