An energy system control method and device, electronic equipment and medium

Abstract

Embodiments of the present application provide an energy system control method and device, electronic equipment and medium, comprising: in response to a regulation instruction for an energy system, acquiring multi-modal data, scheduling data corresponding to the regulation instruction; splicing the multi-modal data and the scheduling data to determine a time sequence feature sequence; performing load time sequence prediction on the time sequence feature sequence to determine scheduling information; analyzing the scheduling information to determine an original action sequence; projecting the original action sequence with physical constraints to determine a target action sequence; and controlling the energy system based on the target action sequence. Through the embodiments of the present application, multi-modal information can be fused for output management, improving control accuracy and control quality. Furthermore, the original action is constrained by physical constraints, so that real-time closed-loop energy management decisions can be made in milliseconds on the premise that all physical safety constraints of the energy system are met, improving the explainability and accuracy of control.

Description

Technical Field

[0001] This invention relates to the field of energy control technology, and in particular to an energy system control method, an energy system control device, an electronic device, and a computer-readable storage medium. Background Technology

[0002] Integrated energy systems, by coupling multiple energy forms such as electricity, heat, and natural gas, can significantly improve energy utilization efficiency and renewable energy absorption capacity. However, energy management and control face challenges: the contradiction between strong uncertainty and real-time requirements; the dramatic fluctuations in renewable energy output, load demand, and market prices; and the long solution times (minutes to hours) of traditional mixed-integer linear programming optimization methods, which cannot meet the millisecond to second-level real-time control requirements of the power grid. The contradiction between physical security and the black box nature of AI (artificial intelligence); while existing methods based on deep learning or reinforcement learning have fast response speeds, they lack physical interpretability, leading to equipment damage or power grid instability and insufficient control accuracy. Furthermore, energy management relies on time-series data, making joint reasoning with structured and unstructured data impossible, resulting in insufficient control accuracy. Summary of the Invention

[0003] In view of the above problems, embodiments of the present invention are proposed to provide an energy system control method, an energy system control device, an electronic device, and a computer-readable storage medium that overcome or at least partially solve the above problems.

[0004] To address the aforementioned problems, in a first aspect, an embodiment of the present invention discloses an energy system control method, comprising: In response to control commands for the energy system, acquire multimodal data and scheduling data corresponding to the control commands; The multimodal data and the scheduling data are spliced ​​together to determine a temporal feature sequence; Load time-series prediction is performed on the time-series feature sequences to determine scheduling information; The scheduling information is analyzed to determine the original action sequence; The original action sequence is physically constrained and projected to determine the target action sequence. The energy system is controlled based on the target action sequence.

[0005] Optionally, the step of projecting the original action sequence with physical constraints to determine the target action sequence includes: Obtain physical constraints; The feasible domain of actions is determined based on the aforementioned physical constraints. The target action sequence is determined by combining the original action sequence with the action feasible region.

[0006] Optionally, the action feasible region has multiple target action expressions, and the step of determining the target action sequence by combining the original action sequence with the action feasible region includes: A target function is constructed based on the squared Euclidean distance between the original action sequence and the target action expression. Solving the objective function determines the target action sequence.

[0007] Optionally, the step of determining the target action sequence by combining the original action sequence with the action feasible region further includes: Differentiate the original action sequence to determine the Jacobian matrix; The Jacobian matrix is ​​backpropagated.

[0008] Optionally, the step of concatenating the multimodal data and the scheduling data to determine the temporal feature sequence includes: Feature extraction is performed on the multimodal data to determine the energy feature vector; Parse the scheduling data to determine the structured parameters; The structured parameters are converted into a demand feature vector; By combining the energy feature vector and the demand feature vector, a time-series feature sequence is determined.

[0009] Optionally, the step of performing load time-series prediction on the time-series feature sequence to determine scheduling information includes: Perform load time-series rolling prediction on the time-series feature sequence to determine the load status; Determine the scheduling information corresponding to the load status.

[0010] Optionally, the method further includes: Output the scheduling information.

[0011] Optionally, the method further includes: Detect the output load status of the energy system; Based on the output load status, determine the gradient update information; The gradient update information is homomorphically encrypted to generate an aggregated gradient; Get the global model; The aggregated gradient is used to update the global model; the global model is used to perform the steps of load time-series prediction on the time-series feature sequence to determine scheduling information, and / or the step of physical constraint projection on the original action sequence to determine the target action sequence.

[0012] Optionally, the global model is pre-trained using an irrelevant meta-learning algorithm.

[0013] Optionally, the method further includes: Determine the target object to be executed in the target action sequence; Based on the execution object, the target action sequence is hashed to determine the hash data; The hash data is aggregated to form on-chain smart contract information.

[0014] In a second aspect, embodiments of the present invention disclose an energy system control device, comprising: The acquisition module is used to acquire multimodal data and scheduling data corresponding to the control instructions for the energy system in response to the control instructions. The splicing module is used to splice the multimodal data and the scheduling data to determine the temporal feature sequence; The prediction module is used to perform load time-series prediction on the time-series feature sequence to determine scheduling information; The parsing module is used to parse the scheduling information and determine the original action sequence; The projection module is used to project the original action sequence with physical constraints to determine the target action sequence; A control module is used to control the energy system based on the target action sequence.

[0015] Optionally, the projection module includes: The `get` submodule is used to obtain physical constraints. The feasible region determination submodule is used to determine the feasible region of the action based on the physical constraints. The projection submodule is used to determine the target action sequence by combining the original action sequence with the action feasible region.

[0016] Optionally, the action feasible domain has multiple target action expressions, and the projection submodule includes: The construction unit is used to construct a target function based on the squared Euclidean distance between the original action sequence and the target action expression. The solving unit is used to solve the objective function to determine the target action sequence.

[0017] Optionally, the projection submodule also includes: The differentiation unit is used to differentiate the original action sequence and determine the Jacobian matrix; The reverse unit is used to backpropagate the Jacobian matrix.

[0018] Optionally, the splicing module includes: The time-series extraction submodule is used to extract features from the multimodal data and determine the energy feature vector; The parsing submodule is used to parse the scheduling data and determine the structured parameters; A conversion module is used to convert the structured parameters into a requirement feature vector; The splicing submodule is used to combine the energy feature vector and the demand feature vector to determine the time-series feature sequence.

[0019] Optionally, the prediction module includes: The rolling submodule is used to perform load time-series rolling prediction on the time-series feature sequence to determine the load status; The scheduling determination submodule is used to determine the scheduling information corresponding to the load status.

[0020] Optionally, the device further includes: The output module is used to output the scheduling information.

[0021] Optionally, the device further includes: The detection module is used to detect the output load status of the energy system; The gradient module is used to determine gradient update information based on the output load state. An encryption module is used to perform homomorphic encryption on the gradient update information to generate an aggregated gradient; The global module is used to obtain the global model; An update module is used to update the global model using the aggregated gradient; the global model is used to perform the steps of performing load time-series prediction on the time-series feature sequence to determine scheduling information, and / or the step of performing physical constraint projection on the original action sequence to determine the target action sequence.

[0022] Optionally, the global model is pre-trained using an irrelevant meta-learning algorithm.

[0023] Optionally, the device further includes: An object determination module is used to determine the execution object of the target action sequence; The hash module is used to perform a hash operation on the target action sequence based on the execution object to determine the hash data; The aggregation module is used to aggregate the hash data to form on-chain smart contract information.

[0024] In a third aspect, an embodiment of the present invention discloses an electronic device including a processor, a memory, and a computer program stored in the memory and capable of running on the processor, wherein the computer program, when executed by the processor, implements the steps of the energy system control method as described above.

[0025] In a fourth aspect, embodiments of the present invention disclose a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the energy system control method as described above.

[0026] The embodiments of the present invention have the following advantages: This invention, in response to a control command for an energy system, acquires multimodal data and scheduling data corresponding to the control command; concatenates the multimodal data and the scheduling data to determine a time-series feature sequence; performs load time-series prediction on the time-series feature sequence to determine scheduling information; parses the scheduling information to determine the original action sequence; projects the original action sequence with physical constraints to determine the target action sequence; and controls the energy system based on the target action sequence. By concatenating multimodal data and scheduling data to determine the time-series feature sequence, and by using the time-series feature sequence to fuse multimodal data and unstructured scheduling data for joint output management, output management can refer to the content of the scheduling command, improving control accuracy and quality. Furthermore, by using physical constraints to constrain the original actions, millisecond-level real-time closed-loop energy management decisions can be achieved while ensuring that the energy system meets all physical safety constraints, improving the interpretability and accuracy of control. Attached Figure Description

[0027] Figure 1 This is a flowchart illustrating the steps of an embodiment of the energy system control method of the present invention; Figure 2 This is a flowchart illustrating the steps of an embodiment of the energy system control method of the present invention; Figure 3 This is a schematic diagram of the architecture of an example of an energy system control method according to the present invention; Figure 4 This is a flowchart of the physical projection steps of an example of an energy system control method according to the present invention; Figure 5 This is a structural block diagram of an embodiment of an energy system control device according to the present invention; Figure 6 This is a structural block diagram of an electronic device provided in an embodiment of the present invention; Figure 7 This is a structural block diagram of a storage medium provided in an embodiment of the present invention. Detailed Implementation

[0028] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0029] While existing deep learning or reinforcement learning-based methods offer fast response times, they lack physical interpretability, leading to equipment damage or grid instability, and making it difficult for dispatchers to trust and adopt AI suggestions. Multimodal data fusion and knowledge silos are also challenges. Energy management relies not only on time-series data but also on unstructured data such as weather forecasts, market transaction announcements, and equipment maintenance work orders. Existing platforms cannot achieve joint reasoning between structured and unstructured data; furthermore, user-side data cannot be shared due to privacy concerns, making global model training difficult. Some AI+energy management platforms exist, but they are mostly limited to single-point applications such as load forecasting and fault diagnosis, lacking a unified platform architecture that can simultaneously address the three major issues of physical safety, real-time response, and multimodal cognition. This invention addresses how to achieve millisecond-level real-time closed-loop energy management decisions while ensuring all physical safety constraints of the integrated energy system, such as power balance, branch flow limits, and thermo-hydraulic coupling, through physically guided deep reinforcement learning, multimodal energy learning, and a federated learning cloud-edge collaborative architecture. This overcomes the contradiction between the slow solution of traditional optimization methods and the lack of security guarantees in existing AI methods.

[0030] One of the core concepts of this invention is that the Physics-Guided Deep Reinforcement Learning (PG-DRL) algorithm addresses the issue that traditional reinforcement learning output actions may violate power grid safety constraints, and traditional optimization solvers are slow. The Multi-Modal Energy Foundation Model (MM-EFM) addresses the limitations of existing platforms in handling unstructured data, lack of semantic understanding, and low efficiency in human-computer interaction. A privacy-preserving cloud-edge collaborative framework based on federated learning addresses the difficulty of global model training and poor generalization ability of single-point models caused by the inability to share user-side data due to privacy concerns.

[0031] Reference Figure 1 The diagram illustrates a flowchart of an embodiment of an energy system control method according to the present invention. The energy system control method may specifically include the following steps: Step 101: In response to the control command for the energy system, acquire multimodal data and scheduling data corresponding to the control command; When energy system managers need to control the energy system, they can issue control commands. For example, energy system managers can input voice information as control commands. Upon receiving a control command for the energy system, the system can respond by acquiring multimodal data and corresponding scheduling data. Multimodal data includes data from multiple dimensions, such as time-series data, runtime text data, and image data. The scheduling data corresponding to the control command can be the natural language text data corresponding to the control command.

[0032] Step 102: Concatenate the multimodal data and the scheduling data to determine the temporal feature sequence; Multimodal data and scheduling data can be concatenated, features aligned, and time-based changes, demand and state trends identified to determine time-series feature sequences.

[0033] Step 103: Perform load time-series prediction on the time-series feature sequence to determine scheduling information; Next, load time-series prediction is performed on the time-series characteristic sequence to predict the load state that the energy system needs to reach within a certain period of time in the future. Based on this load state, the execution actions of the control energy system are determined, and scheduling information is determined.

[0034] Step 104: Parse the scheduling information to determine the original action sequence; The scheduling information can be parsed to determine the actions to be executed and the corresponding control variables. Individual actions and their corresponding control variables are combined to form a single raw action, and all raw actions are combined to determine the raw action sequence.

[0035] Step 105: Perform physical constraint projection on the original action sequence to determine the target action sequence; The original action sequence is projected with physical constraints, and each original action in the original action sequence is physically constrained to determine the corresponding target action. All target actions are then combined to generate the target action sequence. The physical constraints can be determined according to actual conditions, such as satisfying power balance, branch power flow limits, and thermo-hydraulic coupling constraints. This embodiment of the invention does not impose limitations.

[0036] Step 106: Control the energy system based on the target action sequence.

[0037] The energy system is controlled by a sequence of target actions to output energy to the outside.

[0038] This invention, in response to a control command for an energy system, acquires multimodal data and scheduling data corresponding to the control command; concatenates the multimodal data and the scheduling data to determine a time-series feature sequence; performs load time-series prediction on the time-series feature sequence to determine scheduling information; parses the scheduling information to determine the original action sequence; projects the original action sequence with physical constraints to determine the target action sequence; and controls the energy system based on the target action sequence. By concatenating multimodal data and scheduling data to determine the time-series feature sequence, and by using the time-series feature sequence to fuse multimodal data and unstructured scheduling data for joint output management, output management can refer to the content of the scheduling command, improving control accuracy and quality. Furthermore, by using physical constraints to constrain the original actions, millisecond-level real-time closed-loop energy management decisions can be achieved while ensuring that the energy system meets all physical safety constraints, improving the interpretability and accuracy of control.

[0039] Reference Figure 2 The diagram illustrates a flowchart of an embodiment of an energy system control method according to the present invention. The energy system control method may specifically include the following steps: Step 201: In response to the control command for the energy system, acquire multimodal data and scheduling data corresponding to the control command; Upon receiving control commands for the energy system, the system can respond to these commands by acquiring multimodal data and corresponding scheduling data. Multimodal data can include time-series data such as output load and output power; text data such as weather warnings, market rules, and work orders; and image data such as satellite cloud images and infrared thermal images. In addition, data cleaning can be performed on multimodal data. The collected multimodal data is examined to remove outliers and missing values. For outliers, their differences from surrounding data are analyzed, and linear interpolation or the average of historical data is used for correction. For missing values, if there are few missing values, the average of the preceding and following data is used to fill in the gaps; if there are many missing values, the historical average data for that period is considered for filling. Data normalization can also be performed. Because the data units and numerical ranges of different features vary greatly, multimodal data can be normalized. For example, a minimum-to-maximum normalization method can be used to map the data to the interval [0, 1].

[0040] Step 202: Concatenate the multimodal data and the scheduling data to determine the temporal feature sequence; Multimodal data and scheduling data can be spliced ​​together to determine a time-series feature sequence. This time-series feature sequence is then used to align and correlate multimodal data across modes, and adjustments are made based on constraints in the scheduling data.

[0041] In an optional embodiment of the present invention, the step of concatenating the multimodal data and the scheduling data to determine the time-series feature sequence includes: extracting features from the multimodal data to determine an energy feature vector; parsing the scheduling data to determine structured parameters; converting the structured parameters into a demand feature vector; and combining the energy feature vector and the demand feature vector to determine the time-series feature sequence.

[0042] Feature extraction can be performed on multimodal data according to its corresponding data type. For example, time-series data sequences within a time window can be encoded into a series of time-series feature vectors as energy feature vectors. For text data such as weather warnings, market rules, and maintenance work orders, textual semantic feature vectors can be extracted as energy feature vectors. For image data such as satellite cloud images and equipment infrared thermal images, image feature vectors can be extracted as energy feature vectors. User-input scheduling data is parsed to extract time, events, targets, etc., forming structured parameters; these structured parameters are converted into demand feature vectors; and the energy feature vectors and demand feature vectors are combined to determine the time-series feature sequence.

[0043] Step 203: Perform load time-series prediction on the time-series feature sequence to determine scheduling information; The load time sequence is predicted based on the spatiotemporal correlation between the system's operating state and output state, and the predicted load value corresponding to each time step is output to determine the scheduling information.

[0044] In an optional embodiment of the present invention, the step of performing load time-series prediction on the time-series feature sequence to determine scheduling information includes: performing load time-series rolling prediction on the time-series feature sequence to determine the load status; and determining the scheduling information corresponding to the load status.

[0045] At each time step, the corresponding load state can be mapped based on the state in the time-series feature sequence. Predictions are then made on a rolling basis, incrementing the time step, to determine the load state within a certain future timeframe. Using this load state as the target, the actions required to control the energy system are determined, generating scheduling information corresponding to the load state.

[0046] For example, the generation of scheduling information can be achieved by constructing a multimodal energy model. This multimodal energy model can be deployed in the cloud of the energy system. First, a multimodal encoder is designed. For time-series data such as load, output, and electricity price, a one-dimensional convolutional network or Transformer time-series encoder is used to encode the sequence within the time window into a series of time-series feature vectors. For text data such as weather warnings, market rules, and maintenance work orders, a pre-trained language model is used to extract textual semantic feature vectors. For image data such as satellite cloud images and equipment infrared thermal images, a convolutional neural network is used to extract image feature vectors. Second, a cross-modal attention fusion module is constructed. The feature vector sequences of the above multimodal models are concatenated and input into a multi-layer Transformer decoder. The decoder uses a cross-attention mechanism to allow text features to focus on relevant time-series segments (e.g., the word "rainstorm" focuses on the photovoltaic output sequence of the next few hours), and image features to guide the extraction of time-series features (e.g., the direction of cloud movement in satellite cloud images is used to correct photovoltaic forecasts). Through pre-training on large-scale multimodal energy data, the model learns cross-modal alignment and correlation knowledge. Then, an integrated large language model is used as the interaction interface. Users input scheduling instructions in natural language (such as "There will be heavy rain and carbon prices will rise tomorrow morning, please generate the economically optimal scheduling plan"). The large language model first parses the scheduling instructions, extracting structured parameters such as time, events, and objectives; then it concatenates the above data to generate a time-series feature sequence. Finally, it directly calls the multimodal fusion model to perform time-series forecasting, gradually predicting the energy system load forecast and scheduling mode, forming scheduling information.

[0047] Step 204: Parse the scheduling information to determine the original action sequence; It can parse the scheduling information, identify each action in the scheduling method, and determine the original action sequence.

[0048] Step 205: Perform physical constraint projection on the original action sequence to determine the target action sequence; The original action sequence can be physically constrained and projected, and the actions in the original action sequence can be physically constrained and guided so that the final target action can be executed in a reasonable and safe form, thus generating the target action sequence.

[0049] In an optional embodiment of this application, the step of performing physical constraint projection on the original action sequence to determine the target action sequence includes: Sub-step S2051: Obtain physical constraints; It can obtain physical constraints, such as power system flow limits, equipment limits, and thermal balance.

[0050] Sub-step S2052: Determine the feasible domain of actions based on the physical constraints; The corresponding feasible domain for action is determined based on physical constraints, such as power balance value, branch power flow limit, thermo-hydraulic coupling ratio, equipment limit, etc.

[0051] Sub-step S2053: Determine the target action sequence by combining the original action sequence with the action feasible region.

[0052] The original action sequence is mapped to the action feasible region, it is determined whether each action in the original action sequence exceeds the action feasible region, and actions within the action feasible region are identified, thereby determining the target action sequence.

[0053] Specifically, the action feasible region has multiple target action expressions, and the step of determining the target action sequence by combining the original action sequence and the action feasible region includes: constructing an objective function based on the squared Euclidean distance between the original action sequence and the target action expression; and solving the objective function to determine the target action sequence.

[0054] Furthermore, in an optional embodiment of the present invention, the step of determining the target action sequence by combining the original action sequence and the action feasible region further includes: taking the derivative of the original action sequence to determine the Jacobian matrix; and backpropagating the Jacobian matrix.

[0055] Furthermore, the original action sequence can be differentiated by first-order calculation to determine the Jacobian matrix, thereby making the physical constraint projection formula differentiable and facilitating correction of the projection process. The Jacobian matrix can be directly used for backpropagation to perform gradient correction of the projection process.

[0056] The physical constraints of the system, such as the power balance equation, branch power flow linearization limits, equipment output limits, thermal supply and demand balance, and energy storage SOC dynamic range, can all be expressed as linear equations and linear inequalities. These linear equations and inequalities constitute the target action expression. The actions corresponding to the target action expression can be vectorized. The squared Euclidean distance between the vectors corresponding to actions in the original action sequence and the vectors corresponding to actions in the target action expression is used to identify the similarity between the actions in the original action sequence and the actions in the target action expression. The objective function is constructed with the goal of minimizing the squared Euclidean distance. Then, the objective function is solved as a convex quadratic programming problem. The action with the minimum squared Euclidean distance is taken as the target action, and all target actions are combined to form the target action sequence.

[0057] For example, a differentiable physical constraint projection layer can be used to process physical constraint projections by a deep reinforcement learning agent. This physical constraint projection layer can be deployed at the edge of the energy system. The physical constraint projection layer maps the original actions output by the policy network to a feasible region that satisfies physical constraints such as power balance, branch flow limits, and thermo-hydraulic coupling through a projection algorithm. The projection process is differentiable and supports end-to-end training. Simultaneously, a multi-objective hybrid reward function incorporating economic costs, safety penalties, and carbon emissions is designed, enabling the agent to automatically learn safe and economical strategies during exploration. Specifically, firstly, the energy management problem of the integrated energy system is modeled as a Markov decision process, defining a state space (including electrical quantities, thermal quantities, environmental quantities, price signals, etc.), an action space (including continuous actions such as unit output and energy storage power, and discrete actions such as start-stop control), and a multi-objective reward function (including economic costs, carbon emission costs, and safety penalties). Secondly, a deep reinforcement learning policy network is constructed, which outputs an original action vector based on the current state. This original action may violate physical constraints such as power flow equations, equipment limits, and thermal balance. Then, the original action vector is input into a differentiable physical constraint projection layer. The function of this layer is to find a legal action vector that is closest to the original action vector in Euclidean distance, while ensuring all preset physical constraints are met. This projection problem is formulated as a convex quadratic programming problem, with all constraints being linear (including power balance equality constraints, branch power flow inequality constraints, upper and lower limits of equipment output, thermal supply and demand balance, and the dynamic range of energy storage SOC, etc.). Due to the convexity of this problem, the solution exists and is unique. The derivative of the projection layer's output with respect to the input, i.e., the Jacobian matrix, can be analytically calculated or automatically obtained using implicit function theorems or automatic differentiation techniques. This allows the projection layer to participate in end-to-end backpropagation training as a standard neural network module: when the final action execution effect is unsatisfactory, the gradient of the loss function can be backpropagated through the projection layer, guiding the policy network to adjust its original output. After sufficient training, the policy network will gradually learn to directly output legal actions close to those within the feasible region, thus requiring only a very short time to complete the projection during the inference phase. Since the projection process only involves solving a low-dimensional convex quadratic programming problem (typically with no more than 50 constraints and no more than 10 variables), it can be completed in milliseconds on edge computing devices. Furthermore, a large safety penalty term is introduced into the reward function to ensure that even if the agent encounters a projected action during the exploration phase, it will receive a very low reward if there is still a slight deviation from the limit (numerical error), thereby forcing the policy network to learn a conservative policy. The Alternating Direction Multiplier Method (ADMM) is used as the projection solver, and this algorithm can converge to high accuracy within a fixed number of iterations. The Jacobian matrix of the projection layer output with respect to the input is calculated using the implicit function theorem, making the projection layer differentiable. Step 206: Control the energy system based on the target action sequence; The target action in the target action sequence is sent to the field equipment for execution, thereby realizing the control of the energy system.

[0058] Step 207: Output the scheduling information.

[0059] It can convert scheduling information, including scheduling plans, cost estimates, and safety verification results, into natural language for output, and can also attach visual charts to enable natural language interaction. By realizing human-computer natural language interactive energy management, the usage threshold for managers is significantly reduced.

[0060] Furthermore, to ensure a balance between data confidentiality and accuracy, the data can be encrypted before model training and layout. In an optional embodiment of this application, the method further includes: Step S1: Detect the output load status of the energy system; It can detect the actual output of energy, for example, by obtaining the output load status through sensors.

[0061] Step S2: Determine gradient update information based on the output load status; The output load status can be compared with the scheduling information to determine the deviation in the actual execution process, and the deviation can be used to determine the gradient update information.

[0062] Step S3: Homomorphically encrypt the gradient update information to generate an aggregated gradient; Homomorphic encryption can be applied to gradient update information, such as through a federated averaging algorithm. This involves each edge node uploading encrypted gradients, which are then weighted and averaged in the cloud domain before decryption, thus converting the gradient update information into aggregated gradients. This encryption ensures data confidentiality.

[0063] Step S4: Obtain the global model; The global model can be initialized in the cloud, including common parts of the policy network and the base of the multimodal model. The global model can be retrieved from the cloud.

[0064] Step S5, update the global model using the aggregated gradient; the global model is used to perform the steps of load time-series prediction on the time-series feature sequence to determine scheduling information, and / or the step of physical constraint projection on the original action sequence to determine the target action sequence.

[0065] The global model is updated using aggregated gradients to bring it to converge, so as to perform the steps of load time-series prediction on the time-series feature sequence to determine scheduling information, and / or the step of physical constraint projection on the original action sequence to determine the target action sequence.

[0066] In the pre-training phase, the global model can employ an independent meta-learning algorithm, using data from different devices to train the meta-model. Newly connected devices only need to provide one week's worth of local data, and by performing 5 steps of gradient fine-tuning on the meta-model, 90% of the normal convergence performance can be achieved, significantly shortening the cold start time.

[0067] In an optional embodiment of the present invention, the method further includes: Step S6: Determine the execution target of the target action sequence; The target of the sequence of actions can be identified, including but not limited to energy system operators, power grid companies, and regulatory agencies.

[0068] Step S7: Based on the execution object, perform a hash operation on the target action sequence to determine the hash data; Based on the execution object, the target action sequence obtained from each execution object is hashed, such as hashing the timestamp, status hash, action, and security verification result in the target action sequence to determine the hash data.

[0069] Step S8: Aggregate the hash data to form on-chain smart contract information.

[0070] Aggregated hash data forms on-chain smart contract information. During investigations, the authenticity of the data is verified through hash comparison, and the actions are re-verified to ensure they meet constraints, thus facilitating oversight.

[0071] This invention, in response to a control command for an energy system, acquires multimodal data and scheduling data corresponding to the control command; concatenates the multimodal data and the scheduling data to determine a time-series feature sequence; performs load time-series prediction on the time-series feature sequence to determine scheduling information; parses the scheduling information to determine the original action sequence; projects the original action sequence with physical constraints to determine the target action sequence; controls the energy system based on the target action sequence; and outputs the scheduling information. By concatenating multimodal data and scheduling data to determine the time-series feature sequence, and using the time-series feature sequence to fuse multimodal data and unstructured scheduling data for joint output management, the output management can refer to the content of the scheduling command, improving control accuracy and quality. Furthermore, by using physical constraints to constrain the original actions, millisecond-level real-time closed-loop energy management decisions can be achieved while ensuring that the energy system meets all physical safety constraints, improving the interpretability and accuracy of control. By setting physical constraint projections between the original actions and the final action output, millisecond-level real-time closed-loop energy management decisions are achieved under the premise that the energy system meets all physical safety constraints. This improves the interpretability and accuracy of control and solves the dual problems of traditional reinforcement learning action violations and slow traditional optimization solutions.

[0072] To enable those skilled in the art to clearly understand the implementation process of the embodiments of the present invention, the following example is used for illustration: For example, the energy system equipment includes: 10 MW photovoltaic, 5 MW wind power, two 2 MW gas-fired combined heat and power units, 4MW electric boiler, 2 MW / 8 MWh energy storage system, peak electrical load of 15 MW, and peak thermal load of 6 MW.

[0073] Hardware configuration: Please refer to Figure 3 Two training servers (equipped with multiple GPUs) are deployed in the cloud for deep reinforcement learning model training, multimodal large model training, and federated learning aggregation and updates, along with four blockchain node servers for secure data storage. At the edge, one edge smart gateway (with GPU inference capabilities) is deployed on each edge, running a physically guided deep reinforcement learning agent, a local data preprocessing module, and a federated learning client. End-side devices are equipped with smart sensors (electricity meters, heat meters, weather stations) and controllers (energy storage BMS, CHP unit PLC, electric boiler frequency converters), aggregating data to the edge gateway via industrial protocols for real-time scheduling and execution.

[0074] The implementation of the physics-guided deep reinforcement learning module can be referred to... Figure 4 The process involves inputting environmental conditions (such as grid load and photovoltaic power generation), then processing the input into a deep reinforcement learning policy network (outputting the original action vector), followed by a physical constraint projection layer (ensuring power balance is satisfied, etc.), and finally outputting the legal action (an action that conforms to the physical constraints). Specifically, this includes: Step 1: MDP Modeling. Define the state space: including bus voltage, branch power flow, energy storage SOC, photovoltaic / wind power output, heating water supply / return temperature, heat load, irradiance prediction, wind speed, outdoor temperature, real-time electricity price, real-time carbon price, etc., totaling approximately 120 continuous variables. Define the action space: continuous actions include CHP active power output, electric boiler power, energy storage charging and discharging power, and heating network water supply temperature setpoint; discrete actions include CHP start / stop and electric boiler start / stop. Define the reward function: negative correlation of economic costs (electricity purchase cost, gas purchase cost, electricity sales revenue) plus negative correlation of carbon emission costs, minus a safety penalty term (providing a large negative reward when constraints are violated) and an action smoothing term (suppressing frequent fluctuations).

[0075] Step 2: Construct a differentiable physical constraint projection layer. All constraints of the power system, including power balance equations, branch power flow linearization limits, equipment output limits, thermal supply and demand balance, and energy storage SOC dynamic range, are expressed as linear equations and linear inequalities. A convex quadratic programming model is established: the objective function is to minimize the squared Euclidean distance between the original action and the legal action. The Alternating Direction Multiplier Method (ADMM) is used as the projection solver; this algorithm converges to high accuracy within a fixed number of iterations. The Jacobian matrix of the projection layer output with respect to the input is calculated using the implicit function theorem, making the projection layer differentiable.

[0076] Step 3: Training and Deployment. Pre-training is performed using a digital twin simulator in the cloud. Each action output passes through a projection layer. After pre-training, the policy network and projection layer are deployed together to the edge smart gateway. The edge gateway executes the following cyclically every 100 milliseconds: collecting the latest state, forward computation of the policy network to obtain the original action, solving for the legal action using the projection layer, and issuing instructions to the device. In actual testing, the average time for a single decision was 12 milliseconds, meeting real-time requirements, and no physical constraint violations occurred during continuous operation.

[0077] The implementation of a large-scale multimodal energy model can include: Step 1: Model Construction. The temporal encoder uses one-dimensional convolution plus positional encoding; the text encoder uses a pre-trained BERT model; and the image encoder uses ResNet-50. Cross-modal fusion uses a 6-layer Transformer decoder with 8 attention heads per layer.

[0078] Step 2: Pre-training. Collect publicly available time-series data on load, solar power, and wind power (minute-level data spanning several years), meteorological warning texts (hundreds of thousands of entries), satellite cloud image archives, and equipment infrared fault image libraries. Pre-training is performed using tasks such as masked time-series prediction, text-time-series alignment, and cross-modal contrastive learning. The total data volume reaches terabytes, and training is conducted on a GPU cluster for two weeks.

[0079] Step 3: Natural Language Interaction Integration. ChatGLM3-6B is selected as the dialogue agent, and the output of the multimodal fusion model is injected into the large language model through low-rank adaptation. 5000 manually annotated "instruction-parameter-plan" data pairs are constructed for instruction fine-tuning. The scheduler can input natural language instructions through a web interface; the system automatically parses them, calls the multimodal model to generate predictions, calls the reinforcement learning agent to simulate and evaluate, and finally returns a natural language solution along with a security verification report.

[0080] The implementation of federal learning privacy protections includes the following: Step 1: Federated Learning Setup. Initialize the global model in the cloud (including the common part of the policy network and the base of the multimodal model). Multiple campuses (edge ​​nodes) download the global model and each uses local data (one year of historical running data) to fine-tune it locally and calculate the gradient update amount.

[0081] Step 2: Homomorphic Encryption. A homomorphic encryption scheme is used. Each edge node encrypts its gradient updates using the public key distributed in the cloud and uploads the ciphertext. The cloud performs a weighted average of the ciphertext (ciphertext addition), decrypts it using the private key to obtain the aggregated gradient, and updates the global model. After several rounds of communication, the global model converges.

[0082] Step 3: Meta-learning Cold Start. During the pre-training phase, a model-independent meta-learning algorithm is employed, using data from multiple source environments to train the meta-model. Newly connected deployment environments only need one week's worth of local data. Five steps of gradient fine-tuning based on the meta-model are sufficient to achieve 90% of the normal convergence performance, significantly shortening the cold start time.

[0083] Step 4: Secure Evidence Storage. Establish a consortium blockchain with participants including telecom operators, power grid companies, and regulatory agencies. Perform hash calculations on each scheduling decision (timestamp, state hash, action, and security verification result) and write it into an on-chain smart contract. During incident investigations, regulatory agencies can request the park to provide the original data, verify its authenticity through hash comparison, and re-verify whether the actions meet the constraints.

[0084] It should be noted that, for the sake of simplicity, the method embodiments are all described as a series of actions. However, those skilled in the art should understand that the embodiments of the present invention are not limited to the described order of actions, because according to the embodiments of the present invention, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions involved are not necessarily essential to the embodiments of the present invention.

[0085] Reference Figure 5 The diagram illustrates a structural block diagram of an embodiment of an energy system control device according to the present invention. The energy system control device may specifically include the following modules: The acquisition module 501 is used to acquire multimodal data and scheduling data corresponding to the control command in response to the control command for the energy system. The splicing module 502 is used to splice the multimodal data and the scheduling data to determine the temporal feature sequence; Prediction module 503 is used to perform load time-series prediction on the time-series feature sequence to determine scheduling information; Parsing module 504 is used to parse the scheduling information and determine the original action sequence; Projection module 505 is used to perform physical constraint projection on the original action sequence to determine the target action sequence; The control module 506 is used to control the energy system based on the target action sequence.

[0086] Optionally, the projection module 505 includes: The `get` submodule is used to obtain physical constraints. The feasible region determination submodule is used to determine the feasible region of the action based on the physical constraints. The projection submodule is used to determine the target action sequence by combining the original action sequence with the action feasible region.

[0087] Optionally, the action feasible domain has multiple target action expressions, and the projection submodule includes: The construction unit is used to construct a target function based on the squared Euclidean distance between the original action sequence and the target action expression. The solving unit is used to solve the objective function to determine the target action sequence.

[0088] Optionally, the projection submodule also includes: The differentiation unit is used to differentiate the original action sequence and determine the Jacobian matrix; The reverse unit is used to backpropagate the Jacobian matrix.

[0089] Optionally, the splicing module 502 includes: The time-series extraction submodule is used to extract features from the multimodal data and determine the energy feature vector; The parsing submodule is used to parse the scheduling data and determine the structured parameters; A conversion module is used to convert the structured parameters into a requirement feature vector; The splicing submodule is used to combine the energy feature vector and the demand feature vector to determine the time-series feature sequence.

[0090] Optionally, the prediction module 503 includes: The rolling submodule is used to perform load time-series rolling prediction on the time-series feature sequence to determine the load status; The scheduling determination submodule is used to determine the scheduling information corresponding to the load status.

[0091] Optionally, the device further includes: The output module is used to output the scheduling information.

[0092] Optionally, the device further includes: The detection module is used to detect the output load status of the energy system; The gradient module is used to determine gradient update information based on the output load state. An encryption module is used to perform homomorphic encryption on the gradient update information to generate an aggregated gradient; The global module is used to obtain the global model; An update module is used to update the global model using the aggregated gradient; the global model is used to perform the steps of performing load time-series prediction on the time-series feature sequence to determine scheduling information, and / or the step of performing physical constraint projection on the original action sequence to determine the target action sequence.

[0093] Optionally, the global model is pre-trained using an irrelevant meta-learning algorithm.

[0094] Optionally, the device further includes: An object determination module is used to determine the execution object of the target action sequence; The hash module is used to perform a hash operation on the target action sequence based on the execution object to determine the hash data; The aggregation module is used to aggregate the hash data to form on-chain smart contract information.

[0095] This invention, in response to a control command for an energy system, acquires multimodal data and scheduling data corresponding to the control command; concatenates the multimodal data and the scheduling data to determine a time-series feature sequence; performs load time-series prediction on the time-series feature sequence to determine scheduling information; parses the scheduling information to determine the original action sequence; projects the original action sequence with physical constraints to determine the target action sequence; and controls the energy system based on the target action sequence. By concatenating multimodal data and scheduling data to determine the time-series feature sequence, and by using the time-series feature sequence to fuse multimodal data and unstructured scheduling data for joint output management, output management can refer to the content of the scheduling command, improving control accuracy and quality. Furthermore, by using physical constraints to constrain the original actions, millisecond-level real-time closed-loop energy management decisions can be achieved while ensuring that the energy system meets all physical safety constraints, improving the interpretability and accuracy of control.

[0096] As the device embodiment is basically similar to the method embodiment, the description is relatively simple, and relevant parts can be found in the description of the method embodiment.

[0097] Reference Figure 6 This invention also provides an electronic device, including a processor 601 and a memory 602. The memory 602 stores a computer program executable by the processor 601. When the electronic device is in operation, the processor 601 executes the computer program to implement the integrated energy system control method as described in any one of the embodiments of this invention. The energy system control method includes: In response to control commands for the energy system, acquire multimodal data and scheduling data corresponding to the control commands; The multimodal data and the scheduling data are spliced ​​together to determine a temporal feature sequence; Load time-series prediction is performed on the time-series feature sequences to determine scheduling information; The scheduling information is analyzed to determine the original action sequence; The original action sequence is physically constrained and projected to determine the target action sequence. The energy system is controlled based on the target action sequence.

[0098] Optionally, the step of projecting the original action sequence with physical constraints to determine the target action sequence includes: Obtain physical constraints; The feasible domain of actions is determined based on the aforementioned physical constraints. The target action sequence is determined by combining the original action sequence with the action feasible region.

[0099] Optionally, the action feasible region has multiple target action expressions, and the step of determining the target action sequence by combining the original action sequence with the action feasible region includes: A target function is constructed based on the squared Euclidean distance between the original action sequence and the target action expression. Solving the objective function determines the target action sequence.

[0100] Optionally, the step of determining the target action sequence by combining the original action sequence with the action feasible region further includes: Differentiate the original action sequence to determine the Jacobian matrix; The Jacobian matrix is ​​backpropagated.

[0101] Optionally, the step of concatenating the multimodal data and the scheduling data to determine the temporal feature sequence includes: Feature extraction is performed on the multimodal data to determine the energy feature vector; Parse the scheduling data to determine the structured parameters; The structured parameters are converted into a demand feature vector; By combining the energy feature vector and the demand feature vector, a time-series feature sequence is determined.

[0102] Optionally, the step of performing load time-series prediction on the time-series feature sequence to determine scheduling information includes: Perform load time-series rolling prediction on the time-series feature sequence to determine the load status; Determine the scheduling information corresponding to the load status.

[0103] Optionally, the method further includes: Output the scheduling information.

[0104] Optionally, the method further includes: Detect the output load status of the energy system; Based on the output load status, determine the gradient update information; The gradient update information is homomorphically encrypted to generate an aggregated gradient; Get the global model; The aggregated gradient is used to update the global model; the global model is used to perform the steps of load time-series prediction on the time-series feature sequence to determine scheduling information, and / or the step of physical constraint projection on the original action sequence to determine the target action sequence.

[0105] Optionally, the global model is pre-trained using an irrelevant meta-learning algorithm.

[0106] Optionally, the method further includes: Determine the target object to be executed in the target action sequence; Based on the execution object, the target action sequence is hashed to determine the hash data; The hash data is aggregated to form on-chain smart contract information.

[0107] This invention, in response to a control command for an energy system, acquires multimodal data and scheduling data corresponding to the control command; concatenates the multimodal data and the scheduling data to determine a time-series feature sequence; performs load time-series prediction on the time-series feature sequence to determine scheduling information; parses the scheduling information to determine the original action sequence; projects the original action sequence with physical constraints to determine the target action sequence; and controls the energy system based on the target action sequence. By concatenating multimodal data and scheduling data to determine the time-series feature sequence, and by using the time-series feature sequence to fuse multimodal data and unstructured scheduling data for joint output management, output management can refer to the content of the scheduling command, improving control accuracy and quality. Furthermore, by using physical constraints to constrain the original actions, millisecond-level real-time closed-loop energy management decisions can be achieved while ensuring that the energy system meets all physical safety constraints, improving the interpretability and accuracy of control.

[0108] The memory may include random access memory (RAM) or non-volatile memory, such as at least one disk storage device. Optionally, the memory may also be at least one storage device located remotely from the aforementioned processor.

[0109] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0110] Reference Figure 7This invention also provides a computer-readable storage medium 701, on which a computer program is stored. When a processor executes the computer program, it performs the energy system control method as described in any one of the embodiments of this invention. The energy system control method includes: In response to control commands for the energy system, acquire multimodal data and scheduling data corresponding to the control commands; The multimodal data and the scheduling data are spliced ​​together to determine a temporal feature sequence; Load time-series prediction is performed on the time-series feature sequences to determine scheduling information; The scheduling information is analyzed to determine the original action sequence; The original action sequence is physically constrained and projected to determine the target action sequence. The energy system is controlled based on the target action sequence.

[0111] Optionally, the step of projecting the original action sequence with physical constraints to determine the target action sequence includes: Obtain physical constraints; The feasible domain of actions is determined based on the aforementioned physical constraints. The target action sequence is determined by combining the original action sequence with the action feasible region.

[0112] Optionally, the action feasible region has multiple target action expressions, and the step of determining the target action sequence by combining the original action sequence with the action feasible region includes: A target function is constructed based on the squared Euclidean distance between the original action sequence and the target action expression. Solving the objective function determines the target action sequence.

[0113] Optionally, the step of determining the target action sequence by combining the original action sequence with the action feasible region further includes: Differentiate the original action sequence to determine the Jacobian matrix; The Jacobian matrix is ​​backpropagated.

[0114] Optionally, the step of concatenating the multimodal data and the scheduling data to determine the temporal feature sequence includes: Feature extraction is performed on the multimodal data to determine the energy feature vector; Parse the scheduling data to determine the structured parameters; The structured parameters are converted into a demand feature vector; By combining the energy feature vector and the demand feature vector, a time-series feature sequence is determined.

[0115] Optionally, the step of performing load time-series prediction on the time-series feature sequence to determine scheduling information includes: Perform load time-series rolling prediction on the time-series feature sequence to determine the load status; Determine the scheduling information corresponding to the load status.

[0116] Optionally, the method further includes: Output the scheduling information.

[0117] Optionally, the method further includes: Detect the output load status of the energy system; Based on the output load status, determine the gradient update information; The gradient update information is homomorphically encrypted to generate an aggregated gradient; Get the global model; The aggregated gradient is used to update the global model; the global model is used to perform the steps of load time-series prediction on the time-series feature sequence to determine scheduling information, and / or the step of physical constraint projection on the original action sequence to determine the target action sequence.

[0118] Optionally, the global model is pre-trained using an irrelevant meta-learning algorithm.

[0119] Optionally, the method further includes: Determine the target object to be executed in the target action sequence; Based on the execution object, the target action sequence is hashed to determine the hash data; The hash data is aggregated to form on-chain smart contract information.

[0120] This invention, in response to a control command for an energy system, acquires multimodal data and scheduling data corresponding to the control command; concatenates the multimodal data and the scheduling data to determine a time-series feature sequence; performs load time-series prediction on the time-series feature sequence to determine scheduling information; parses the scheduling information to determine the original action sequence; projects the original action sequence with physical constraints to determine the target action sequence; and controls the energy system based on the target action sequence. By concatenating multimodal data and scheduling data to determine the time-series feature sequence, and by using the time-series feature sequence to fuse multimodal data and unstructured scheduling data for joint output management, output management can refer to the content of the scheduling command, improving control accuracy and quality. Furthermore, by using physical constraints to constrain the original actions, millisecond-level real-time closed-loop energy management decisions can be achieved while ensuring that the energy system meets all physical safety constraints, improving the interpretability and accuracy of control.

[0121] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0122] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, apparatus, or computer program products. Therefore, embodiments of the present invention can take the form of entirely hardware embodiments, entirely software embodiments, or embodiments combining software and hardware aspects. Furthermore, embodiments of the present invention can take the form of computer program products implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0123] Embodiments of the present invention are described with reference to flowchart illustrations and / or block diagrams of methods, terminal devices (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing terminal device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing terminal device, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0124] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing terminal device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0125] These computer program instructions can also be loaded onto a computer or other programmable data processing terminal equipment, causing a series of operational steps to be performed on the computer or other programmable terminal equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable terminal equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0126] Although preferred embodiments of the present invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the embodiments of the present invention.

[0127] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal device. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal device that includes said element.

[0128] The present invention has provided a detailed description of an energy system control method, an energy system control device, an electronic device, and a computer-readable storage medium. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. An energy system control method, characterized in that, include: In response to control commands for the energy system, acquire multimodal data and scheduling data corresponding to the control commands; The multimodal data and the scheduling data are spliced ​​together to determine a temporal feature sequence; Load time-series prediction is performed on the time-series feature sequences to determine scheduling information; The scheduling information is analyzed to determine the original action sequence; The original action sequence is physically constrained and projected to determine the target action sequence. The energy system is controlled based on the target action sequence.

2. The method according to claim 1, characterized in that, The step of performing a physical constraint projection on the original action sequence to determine the target action sequence includes: Obtain physical constraints; The feasible domain of actions is determined based on the aforementioned physical constraints. The target action sequence is determined by combining the original action sequence with the action feasible region.

3. The method according to claim 2, characterized in that, The action feasible region has multiple target action expressions, and the step of determining the target action sequence by combining the original action sequence and the action feasible region includes: A target function is constructed based on the squared Euclidean distance between the original action sequence and the target action expression. Solving the objective function determines the target action sequence.

4. The method according to claim 3, characterized in that, The step of determining the target action sequence by combining the original action sequence with the action feasible region further includes: Differentiate the original action sequence to determine the Jacobian matrix; The Jacobian matrix is ​​backpropagated.

5. The method according to claim 1, characterized in that, The step of concatenating the multimodal data and the scheduling data to determine the temporal feature sequence includes: Feature extraction is performed on the multimodal data to determine the energy feature vector; Parse the scheduling data to determine the structured parameters; The structured parameters are converted into a demand feature vector; By combining the energy feature vector and the demand feature vector, a time-series feature sequence is determined.

6. The method according to claim 1, characterized in that, The step of performing load time-series prediction on the time-series feature sequence to determine scheduling information includes: Perform load time-series rolling prediction on the time-series feature sequence to determine the load status; Determine the scheduling information corresponding to the load status.

7. The method according to claim 1, characterized in that, The method further includes: Output the scheduling information.

8. The method according to claim 1, characterized in that, The method further includes: Detect the output load status of the energy system; Based on the output load status, determine the gradient update information; The gradient update information is homomorphically encrypted to generate an aggregated gradient; Get the global model; The aggregated gradient is used to update the global model; the global model is used to perform the steps of load time-series prediction on the time-series feature sequence to determine scheduling information, and / or the step of physical constraint projection on the original action sequence to determine the target action sequence.

9. The method according to claim 8, characterized in that, The global model is pre-trained using an irrelevant meta-learning algorithm.

10. The method according to any one of claims 1-9, characterized in that, The method further includes: Determine the target object to be executed in the target action sequence; Based on the execution object, the target action sequence is hashed to determine the hash data; The hash data is aggregated to form on-chain smart contract information.

11. An energy system control device, characterized in that, include: The acquisition module is used to acquire multimodal data and scheduling data corresponding to the control instructions for the energy system in response to the control instructions. The splicing module is used to splice the multimodal data and the scheduling data to determine the temporal feature sequence; The prediction module is used to perform load time-series prediction on the time-series feature sequence to determine scheduling information; The parsing module is used to parse the scheduling information and determine the original action sequence; The projection module is used to project the original action sequence with physical constraints to determine the target action sequence; A control module is used to control the energy system based on the target action sequence.

12. An electronic device, characterized in that, It includes a processor, a memory, and a computer program stored in the memory and capable of running on the processor, wherein the computer program, when executed by the processor, implements the steps of the energy system control method as described in any one of claims 1-10.

13. A computer-readable storage medium, characterized in that, A computer program is stored on the computer-readable storage medium, which, when executed by a processor, implements the steps of the energy system control method as described in any one of claims 1-10.

Images