Multi-agent collaborative simulation control system of information energy system

Abstract

The embodiment of the invention provides a multi-agent co-simulation control system of an information energy system, and the system consists of multiple agents, a data processing module, a software defined network and a co-control simulation platform. Unified perception and intelligent decision making of the global state of the energy system are realized, and the self-adaption and autonomy level of the system is improved. Flexible scheduling and safe collaboration of communication resources are realized through the software defined network, and the robustness of the system in dynamic change is enhanced; local control real-time performance and global optimization economical efficiency are effectively balanced through a hierarchical collaborative system of edge, fog and cloud computing; through a high-fidelity closed-loop simulation environment, full-period verification and continuous optimization capability is provided for a control strategy and system evolution.

Description

Technical Field

[0001] The embodiments in this specification relate to the field of simulation technology, and in particular to a multi-agent collaborative simulation control system for information and energy systems. Background Technology

[0002] With the large-scale integration of intermittent distributed energy sources such as wind power and photovoltaics, and the widespread adoption of diversified energy-consuming units such as electric vehicles and flexible loads, modern energy systems are evolving into complex mega-systems with deep coupling of information and energy, a massive number of nodes, and dynamically changing topologies. Existing systems face challenges at the data level: difficulties in integrating multi-source heterogeneous systems, leading to incomplete state awareness; at the control level, centralized or decentralized strategies lack adaptive capabilities, making it difficult to cope with uncertainties and sudden failures; at the coordination level, information networks and physical systems operate independently, lacking cross-domain security and optimization mechanisms; and at the simulation level, existing platforms struggle to perform high-fidelity modeling of the dynamics of cyber-physical coupling and agent interaction, limiting the verification and evolution of new algorithms and strategies.

[0003] Therefore, a better solution is urgently needed. Summary of the Invention

[0004] In view of this, embodiments of this specification provide a multi-agent cooperative simulation control system for information energy systems to address the technical deficiencies existing in the prior art.

[0005] According to a first aspect of the embodiments of this specification, a multi-agent cooperative simulation control system for an information energy system is provided, comprising: Multiple intelligent agents are deployed in the schedulable units of the information and energy system to collect multimodal raw data from the schedulable units and perform local collaborative decision-making and control. The data processing module connects to the intelligent agent and is used to perform interpretable dimensionality reduction processing on the multimodal raw data to obtain dimensionality-reduced data, and to perform feature dimensionality enhancement processing on the dimensionality-reduced data to obtain dimensionality-enhanced feature data. Software-defined networks connect intelligent agents and data processing modules, providing dynamic resource scheduling and secure communication for interactions between intelligent agents. The collaborative control simulation platform connects intelligent agents, data processing modules, and software-defined networks. It is used to integrate edge computing nodes, fog computing nodes, and cloud computing nodes to form a hierarchical control system and to perform closed-loop simulation of systems containing intelligent agents, data processing modules, and software-defined networks.

[0006] In one possible implementation, the data processing module includes a stacked autoencoder and a radial basis function network. The stacked autoencoder is used to reduce the dimensionality of the original multimodal data, and the radial basis function network is used to perform nonlinear mapping on the reduced data to obtain dimensionality-enhanced feature data.

[0007] In one possible implementation, the intelligent agent includes a perception module, a fusion cognition module, and a decision execution module. The perception module is used to collect electrical quantities, physical quantities, state quantities, and environmental quantities. The fusion cognition module is used to extract and fuse features from the collected data. The decision execution module is used to make local decisions and perform collaborative control based on the fused cognition.

[0008] In one possible implementation, software-defined networking adopts a control-forward separation architecture, including a control plane and a data plane. The control plane is equipped with network controllers to implement a global network view and resource scheduling, while the data plane consists of communication interfaces between distributed network devices and agents.

[0009] In one possible implementation, the collaborative management simulation platform includes an edge computing layer, a fog computing layer, and a cloud computing layer. The edge computing layer is used to enable rapid local control of intelligent agents, the fog computing layer is used to coordinate multiple intelligent agents for local optimization, and the cloud computing layer is used for data analysis and model training.

[0010] In one possible implementation, the collaborative management and control simulation platform has a built-in simulation engine that simulates a network topology containing energy nodes and supports setting up running scenarios to test different control strategies.

[0011] In one possible implementation, the stacked autoencoder includes an encoding network and a decoding network, the decoding network being used to reconstruct the dimensionality-reduced data and preserve interpretable features by comparing the reconstructed data with the original multimodal data.

[0012] In one possible implementation, the decision execution module employs reinforcement learning algorithms, fuzzy logic algorithms, or rule base algorithms for local collaborative decision control.

[0013] In one possible implementation, the network controller is also used to implement access control policies, traffic monitoring, and intrusion detection.

[0014] In one possible implementation, a network swarm optimization layer is set up between the fog computing layer and the cloud computing layer. The network swarm optimization layer uses swarm intelligence optimization algorithms to optimize network resource allocation.

[0015] This specification provides a multi-agent collaborative simulation control system for information energy systems. This system, by constructing a system composed of multiple agents, a data processing module, a software-defined network, and a collaborative management and control simulation platform, achieves unified perception and intelligent decision-making of the entire energy system's state, enhancing the system's adaptability and autonomy. Through the software-defined network, it enables flexible scheduling and secure collaboration of communication resources, strengthening the system's robustness in dynamic changes. A layered collaborative system utilizing edge computing, fog computing, and cloud computing effectively balances the real-time performance of local control with the economic efficiency of global optimization. A high-fidelity closed-loop simulation environment provides full-cycle verification and continuous optimization capabilities for control strategies and system evolution. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of a multi-agent cooperative simulation control system for an information energy system provided in one embodiment of this specification. Detailed Implementation

[0017] Many specific details are set forth in the following description to provide a full understanding of this specification. However, this specification can be implemented in many other ways than those described herein, and those skilled in the art can make similar extensions without departing from the spirit of this specification. Therefore, this specification is not limited to the specific implementations disclosed below.

[0018] The terminology used in one or more embodiments of this specification is for the purpose of describing particular embodiments only and is not intended to be limiting of the one or more embodiments of this specification. The singular forms “a” and “the” as used in one or more embodiments of this specification and the appended claims are also intended to include the plural forms, unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in one or more embodiments of this specification refers to and includes any or all possible combinations of one or more associated listed items.

[0019] It should be understood that although the terms first, second, etc., may be used to describe various information in one or more embodiments of this specification, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, first may also be referred to as second without departing from the scope of one or more embodiments of this specification, and similarly, second may also be referred to as first. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to a determination."

[0020] This specification provides a multi-agent cooperative simulation control system for information energy systems, which will be described in detail in the following embodiments.

[0021] See Figure 1 , Figure 1 This document illustrates a schematic diagram of a multi-agent collaborative simulation control system for an information energy system according to an embodiment of this specification. Specifically, it includes: multiple agents deployed in a schedulable unit of the information energy system, used to collect multimodal raw data from the schedulable unit and perform local collaborative decision-making and control; a data processing module connected to the agents, used to perform interpretable dimensionality reduction processing on the multimodal raw data to obtain dimensionality-reduced data, and to perform feature augmentation processing on the dimensionality-reduced data to obtain dimensionality-enhanced feature data; a software-defined network connecting the agents and the data processing module, used to provide dynamic resource scheduling and secure communication for interactions between agents; and a collaborative management and control simulation platform connecting the agents, the data processing module, and the software-defined network, used to integrate edge computing nodes, fog computing nodes, and cloud computing nodes to form a hierarchical control system, and to perform closed-loop simulation of the system containing the agents, the data processing module, and the software-defined network.

[0022] Among these, "information energy system" refers to a complex system deeply coupled with information networks and physical energy networks (such as power grids and heating networks). "Multi-agent" refers to abstracting dispatchable units in a system, such as photovoltaic inverters, energy storage converters, and smart loads, into software entities with autonomous decision-making capabilities. "Multimodal raw data" refers to operational data collected by agents in various formats and from diverse sources, including electrical quantities (voltage, current), physical quantities (temperature), state quantities (equipment switching states), and environmental quantities (light intensity). "Interpretable dimensionality reduction" refers to the process of reducing data dimensionality through algorithms while preserving the key physical meaning of the data. "Feature dimensionality enhancement" refers to the process of generating new features with stronger representational capabilities by mining deep patterns in the data. "Software-defined network" refers to a network architecture that separates the network control plane from the data forwarding plane and allows for programmable management through a central controller. "Edge computing node" refers to a computing unit deployed near physical devices, responsible for rapid local data processing and control. "Fog computing node" refers to a computing unit deployed at a regional convergence point, responsible for collaborative optimization of multiple agents within a certain range. A cloud computing node can refer to a computing unit deployed in a remote data center, responsible for large-scale data storage, analysis, and model training. Closed-loop simulation can refer to a complete simulation of the continuous process from data acquisition, agent decision-making, network communication to system feedback in a virtual environment.

[0023] The present invention will be further described below through a detailed embodiment.

[0024] In a regional integrated energy system scenario that includes multiple commercial buildings, photovoltaic power stations and energy storage stations, a multi-agent collaborative simulation control system for the aforementioned information energy system is deployed.

[0025] The system's multiple intelligent agents are embedded in the energy management systems of various buildings, the controllers of photovoltaic inverters, and the battery management units of energy storage systems. These intelligent agents operate continuously, collecting multimodal raw data from their respective schedulable units through their integrated sensors and data interfaces. This data includes things like the building's real-time power consumption, the output current and voltage of the photovoltaic panels, the state of charge of the energy storage batteries, and local weather station temperature data.

[0026] The collected multimodal raw data is transmitted in real time to the cloud-based data processing module via the data plane of a software-defined network. This module first cleans and standardizes the input raw data to eliminate the influence of dimensions and noise. Then, interpretable dimensionality reduction is performed: a pre-trained stacked autoencoder network compresses the high-dimensional raw data into low-dimensional data containing the core operational state. This process extracts abstract features through the encoder and reconstructs the data through the decoder to verify whether key features are preserved, ensuring the interpretability of the dimensionality reduction results. For example, the dimensionality-reduced vectors clearly reflect the contradictory operational state of "high power output and low illumination." Next, the dimensionality-reduced data is fed into a radial basis function network for feature augmentation. This network, through its nonlinear mapping capabilities, derives more discriminative augmented feature data from the dimensionality-reduced data, such as generating feature vectors for predicting potential equipment failure risks or short-term load changes.

[0027] The generated augmented feature data is fed back to the collaborative management and control simulation platform and distributed to the relevant agents. The control plane (i.e., the network controller) of the software-defined network dynamically calculates the optimal communication path based on the real-time needs of agents to exchange augmented feature data or collaborative commands, and reserves bandwidth for high-priority control commands to ensure low latency and high reliability of the interaction.

[0028] The collaborative management and control simulation platform coordinates the operation of the entire hierarchical control system. Intelligent agents (edge ​​computing nodes) located within buildings utilize received augmented feature data to make local decisions within milliseconds, such as adjusting the set temperature of an air conditioner within the building. Fog computing nodes located in the regional energy management center aggregate augmented feature data from all buildings and distributed power sources within their jurisdiction, perform minute-level optimization calculations, formulate the most economical power procurement and allocation plan, and distribute it to each intelligent agent for execution. In the cloud, cloud computing nodes use historical and real-time aggregated full augmented feature data to train deep neural network models for load forecasting and fault diagnosis, and periodically distribute updated model parameters, enabling edge and fog computing nodes to continuously evolve their decision-making capabilities.

[0029] Furthermore, the entire system can run in a digital twin mode within the simulation engine built into the collaborative management and control simulation platform. Users can construct a simulation model corresponding to the physical area, containing tens of thousands of virtual nodes, and set up extreme scenarios such as "large-scale photovoltaic grid disconnection." The simulation engine will simulate the complete closed loop of data acquisition by the intelligent agent, feature extraction by the processing module, communication of the software-defined network, and collaborative decision-making among the computing nodes, thereby verifying the effectiveness and robustness of different control strategies before practical application.

[0030] The beneficial effects of this embodiment are that by integrating intelligent agents, data processing, software-defined networks, and a hierarchical collaborative simulation platform into a unified system, it achieves digital modeling and intelligent control of all elements of the information energy system. This system can transform multi-source heterogeneous raw data into interpretable, high-quality features, empowering intelligent agents at all levels to make accurate decisions; the software-defined network ensures the flexibility and security of collaborative communication among massive intelligent agents; the "edge-fog-cloud" collaboration achieves the scientific decomposition and global optimization of control responsibilities; and the high-fidelity closed-loop simulation capability provides a safe and efficient virtual experimental environment for the verification and evolution of system strategies, thereby comprehensively improving the operating efficiency, autonomy, and risk resistance level of the energy system.

[0031] In one possible implementation, the data processing module includes a stacked autoencoder and a radial basis function network. The stacked autoencoder is used to reduce the dimensionality of the original multimodal data, and the radial basis function network is used to perform nonlinear mapping on the reduced data to obtain dimensionality-enhanced feature data.

[0032] Stacked autoencoders can refer to a type of neural network consisting of multiple encoder and decoder layers stacked together, used to learn hierarchical feature representations of data. Radial basis function networks can refer to a feedforward neural network that uses radial basis functions as activation functions, excelling at function approximation and pattern classification.

[0033] The specific implementation method of the data processing module is as follows.

[0034] The stacked autoencoder consists of an encoder and a decoder. The encoder comprises three fully connected layers that sequentially map the original multimodal input data (e.g., a vector containing multiple parameters such as voltage, current, power, and temperature) to a hidden space with progressively decreasing dimensionality. Each encoder layer employs a denoising training strategy during pre-training, adding random noise to the input data and training the layer to reconstruct a clean input. This forces the network to learn noise-resistant features. After all layers are pre-trained, overall fine-tuning is performed. The decoder is symmetrical to the encoder and consists of deconvolution and unpooling operations. It reconstructs the dimensionality-reduced data (i.e., the output of the smallest encoder layer) into data with the same dimensionality as the original input. By calculating the mean squared error between the reconstructed and original data and backpropagating to optimize the network parameters, the dimensionality reduction process ensures that the key structures and interpretable information in the original data are preserved to the greatest extent possible, rather than simply compressing the data.

[0035] The dimensionality-reduced data, obtained through stacked autoencoder dimensionality reduction, is then fed into a radial basis function network. This network consists of an input layer, a hidden layer with a Gaussian kernel, and an output layer. Each neuron in the hidden layer corresponds to a "center point," and its output is obtained by transforming the Euclidean distance between the input data and the neuron's center point using a Gaussian function. The network is trained using supervised learning with high-level application tasks (such as fault classification labeling) as the target, adjusting the positions of the hidden layer center points and the weights of the output layer. After training, the network can non-linearly map the input dimensionality-reduced data to a higher-dimensional or more discriminative feature space, thus obtaining dimensionality-enhanced feature data. This data not only contains compressed information from the original data but also incorporates deep patterns related to specific high-level tasks, providing higher-quality input for subsequent intelligent decision-making.

[0036] The beneficial effect of this embodiment lies in the fact that by employing a cascaded structure of a stacked autoencoder and a radial basis function network, a complete processing chain is formed, from data cleaning and interpretable compression to deep feature mining. The stacked autoencoder ensures the interpretability and information fidelity of the data dimensionality reduction process, avoiding "black box" operations; the radial basis function network, with its powerful nonlinear fitting capability, further improves the representation quality of features. The combination of the two enables the data processing module to provide upper-layer applications with concise yet informative features, which is the cornerstone for improving the overall system intelligence level.

[0037] In one possible implementation, the intelligent agent includes a perception module, a fusion cognition module, and a decision execution module. The perception module is used to collect electrical quantities, physical quantities, state quantities, and environmental quantities. The fusion cognition module is used to extract and fuse features from the collected data. The decision execution module is used to make local decisions and perform collaborative control based on the fused cognition.

[0038] The perception module refers to the functional unit within an intelligent agent responsible for acquiring raw information from the physical world or data interfaces. The fusion cognition module refers to the functional unit responsible for integrating, analyzing, and understanding multi-source data to form a comprehensive judgment of itself and its surrounding environment. The decision execution module refers to the functional unit that, based on current cognition, selects and implements specific control strategies and can collaborate with other intelligent agents.

[0039] Specifically, let's take an intelligent agent deployed in a home energy storage converter as an example to illustrate its internal structure.

[0040] The sensing module of this intelligent agent integrates voltage / current sensors, a communication interface for the battery management system, and a data bus for connecting to the home energy management system. It can periodically collect electrical quantities (such as AC side voltage and output power), physical quantities (such as inverter radiator temperature), state quantities (such as battery state of charge and inverter operating mode), and environmental quantities (such as real-time electricity price signals obtained from the home gateway).

[0041] The collected multi-source data is fed into the fusion and cognition module. This module incorporates a lightweight temporal convolutional network model to extract transient and steady-state features from the time-series data of voltage and current. Simultaneously, an attention mechanism submodule is used to perform weighted fusion of multi-dimensional data such as temperature, state of charge, and electricity price to assess the comprehensive status of the current "equipment health," "economic efficiency," and "grid support requirements," forming a structured understanding of the current operating situation of the energy storage unit.

[0042] The decision execution module receives structured cognition from the fusion cognition module. This module maintains a rule base containing various control strategies (such as "peak-valley arbitrage," "power smoothing," and "demand response") and integrates a lightweight reinforcement learning model based on Q-learning. Based on the current cognition (e.g., "electricity prices are at their peak and battery health is good"), the decision execution module makes local decisions, such as "discharging to the grid at rated power," by querying the rule base or running the reinforcement learning model. Simultaneously, this module receives collaborative requests and information from neighboring photovoltaic agents or higher-level fog computing nodes via a software-defined network. Upon receiving a collaborative instruction of "regional voltage exceeding limits, requesting reactive power support," the decision execution module integrates local objectives with collaborative needs, adjusts its control strategy (e.g., injecting a certain amount of reactive power while discharging), achieves collaborative control with other agents, and feeds back the execution results through the communication interface.

[0043] The beneficial effect of this embodiment lies in that by designing a complete functional closed loop for the intelligent agent, including perception, fusion cognition, and decision execution, each energy node is endowed with anthropomorphic "senses," "brain," and "hands and feet." This enables the intelligent agent not only to fully perceive its own state and environment but also to understand complex situations and make autonomous or collaborative decisions that take into account both local goals and global interests. This fundamentally changes the traditional passive and controlled mode of energy equipment and is the foundation for realizing distributed swarm intelligence.

[0044] In one possible implementation, software-defined networking adopts a control-forward separation architecture, including a control plane and a data plane. The control plane is equipped with network controllers to implement a global network view and resource scheduling, while the data plane consists of communication interfaces between distributed network devices and agents.

[0045] In this context, the control plane can refer to the logically centralized control layer responsible for generating and managing network forwarding rules and maintaining the overall network state. The data plane can refer to the set of network devices responsible for executing actual operations such as packet forwarding and processing based on the rules issued by the control plane. The global network view can refer to the complete picture of the entire network topology, link status, traffic load, and other real-time information held by the network controller.

[0046] Based on the above system implementation examples, the specific deployment of software-defined networks is as follows.

[0047] The control plane is controlled by a high-reliability server deployed in a regional data center, which acts as the network controller. This controller runs customized software based on an open-source platform (such as ONOS). It establishes secure connections with all data plane devices through southbound interface protocols (such as OpenFlow), periodically collecting bandwidth utilization, latency, and packet loss rate information for each link, thereby constructing and updating a global network view in real time.

[0048] The data plane includes core and aggregation layer energy switches deployed in power distribution rooms, access layer energy routers deployed in buildings, and communication interface chips supporting the OpenFlow protocol integrated within each intelligent agent. These devices constitute a layered network forwarding infrastructure.

[0049] The workflow is as follows: When an agent (e.g., A) needs to send a collaborative optimization command to another agent (e.g., B), A's communication interface sends the data packet to the energy router it is connected to. The energy router queries its local flow table; if no matching forwarding rule is found for this communication, it encapsulates the first copy of the data packet and uploads it to the network controller via a secure channel. Based on its global network view, the network controller dynamically calculates an optimal path from A to B, considering the current network load and real-time communication requirements, and distributes the corresponding flow table rules to all energy routers and switches along the path. Subsequently, subsequent data packets between A and B are forwarded at high speed in the data plane according to this rule, without further processing by the controller, thus achieving efficient and programmable communication scheduling.

[0050] The beneficial effect of this embodiment is that by adopting a software-defined network architecture that separates control and forwarding, the intelligence of the network is centralized in the controller, while the forwarding devices remain simple and efficient. This enables the network to perform flexible and dynamic resource scheduling and path optimization according to the needs of energy services (such as low-latency control commands and high-bandwidth state synchronization), effectively solving the problem of traditional networks being rigid and unable to adapt to the dynamic collaborative needs of information and energy systems, and providing a solid communication foundation for reliable interaction between large-scale intelligent agents.

[0051] In one possible implementation, the collaborative management simulation platform includes an edge computing layer, a fog computing layer, and a cloud computing layer. The edge computing layer is used to enable rapid local control of intelligent agents, the fog computing layer is used to coordinate multiple intelligent agents for local optimization, and the cloud computing layer is used for data analysis and model training.

[0052] The edge computing layer can refer to computing capabilities deployed near physical devices, using the aforementioned intelligent agents as carriers, responsible for millisecond to second-level rapid response. The fog computing layer can refer to computing units deployed at regional aggregation nodes (such as power distribution room controllers or microgrid central controllers), responsible for minute-level regional coordination and optimization. The cloud computing layer can refer to computing resource pools deployed in remote data centers, responsible for long-term, large-scale data analysis, model training, and macro-level strategy formulation.

[0053] Based on the system's implementation examples, the collaborative management simulation platform's three-tier architecture collaborative working mode is as follows.

[0054] The edge computing layer refers to the intelligent agents themselves deployed in fields such as photovoltaic inverters and energy storage devices. They utilize built-in computing resources to perform local control with extremely high real-time requirements. For example, if a photovoltaic intelligent agent detects a momentary drop in grid voltage, its decision-making and execution module can autonomously decide and execute a control strategy of reducing active power output and injecting reactive current within 100 milliseconds, achieving rapid local "self-healing." This is a function of the edge computing layer.

[0055] The fog computing layer consists of servers (fog computing nodes) deployed in the community's power distribution room or microgrid central controller. These nodes, through a software-defined network, aggregate augmented feature data and status information uploaded by all agents within their jurisdiction (e.g., 50 residential solar panels, 20 energy storage systems, and 10 charging piles). Every 15 minutes, the fog computing node runs a local optimal power flow algorithm, aiming to minimize network losses or achieve the most stable voltage. It coordinates and calculates the active / reactive power output reference values ​​for each agent and distributes the results to the corresponding agents for execution. For example, coordinating the charging or discharging of several energy storage agents during specific time periods can smooth out regional net load fluctuations, achieving localized collaborative optimization.

[0056] The cloud computing layer is deployed in remote cloud data centers and consists of high-performance server clusters. It stores massive amounts of historical and real-time system data and runs complex big data analytics algorithms and deep learning training frameworks. For example, the cloud computing layer uses weather, load, and operational data from the past year to train a long short-term memory network model to predict regional photovoltaic output for the next week. After training, the model is lightweighted and distributed to relevant fog computing nodes and photovoltaic agents to improve their prediction and decision-making capabilities. The cloud also performs cross-regional, long-term strategy mining and economic analysis to formulate a macro-level operational strategy framework.

[0057] The beneficial effect of this embodiment lies in the fact that by clearly defining the responsibilities of the edge, fog, and cloud computing roles, a three-dimensional collaborative management and control system of "local rapid response - regional collaborative optimization - global intelligent evolution" is constructed. This system not only fully utilizes the real-time advantages of edge devices to ensure basic system security, but also achieves efficient integration and optimization of local resources through fog computing; and finally, through the powerful computing power and data insights of the cloud, it drives the continuous iteration and upgrading of the entire system's strategies and models, achieving a unified improvement in control effectiveness, economy, and intelligence level.

[0058] In one possible implementation, the collaborative management and control simulation platform has a built-in simulation engine that simulates a network topology containing energy nodes and supports setting up running scenarios to test different control strategies.

[0059] The simulation engine can refer to the high-fidelity computing core in the aforementioned collaborative management simulation platform, responsible for simulating the dynamic behavior of the information energy system. The network topology of energy nodes can refer to the graph structure composed of model nodes such as transformers, lines, switches, power supplies, and loads, as well as the electrical connections and communication links between them. The operating scenario can refer to the combination of the system's initial state, external disturbances (such as faults or load changes), and boundary conditions (such as weather or electricity prices) preset for testing purposes.

[0060] Based on the system implementation, one of the core functions of this collaborative management and control simulation platform—simulation verification—is achieved through its built-in simulation engine.

[0061] This simulation engine is a software core that employs event-driven parallel discrete event simulation technology. Users can flexibly define a simulated information and energy system through a graphical interface or scripts. For example, they can construct a virtual urban power distribution network containing 20,000 simulation nodes, including transformers, lines, switches, and tens of thousands of photovoltaic, energy storage, electric vehicle charging piles, and smart building load models, and set the electrical connections and communication network topology between them.

[0062] Users can set various operating scenarios in this simulation model. For example, a "summer afternoon thunderstorm" scenario can be set: at 14:00 during the simulation, the output power of 30% of the photovoltaic nodes is suddenly set to zero to simulate a large-scale photovoltaic grid disconnection caused by thunderstorm cloud cover; at the same time, the air conditioning load is set to increase rapidly due to the rise in temperature. Subsequently, users can choose different control strategies to implement in the simulation.

[0063] Once the simulation engine starts running, it accurately simulates the data acquisition, communication interaction, decision-making processes of each intelligent agent and computing nodes at each layer, as well as the dynamic power flow changes of the power network. The engine records and outputs key system metrics throughout the simulation, such as the number of voltage-over-limit nodes, load loss, system recovery time, total network loss, and average delay of control commands. By comparing the metrics under two different strategies, the effectiveness and superiority of the system described in this invention in dealing with extreme scenarios can be quantitatively evaluated.

[0064] The beneficial effect of this embodiment is that, by building a high-fidelity simulation engine, the aforementioned intelligent agents, networks, and collaborative management and control architecture can be tested in a fully controllable and risk-free virtual environment. It can simulate ultra-large-scale complex scenarios, supporting repeated and rapid verification and comparison of innovative algorithms and control strategies, greatly reducing the risks and costs of applying new technologies in real-world systems, and accelerating the research and development and maturation of intelligent energy system solutions.

[0065] In one possible implementation, the stacked autoencoder includes an encoding network and a decoding network, the decoding network being used to reconstruct the dimensionality-reduced data and preserve interpretable features by comparing the reconstructed data with the original multimodal data.

[0066] In this context, the encoding network can refer to the multi-layer neural network structure in the aforementioned stacked autoencoder responsible for converting the input data into a low-dimensional representation. The decoding network can refer to the multi-layer neural network structure in the aforementioned stacked autoencoder responsible for reconstructing the low-dimensional representation back into the original data space. Reconstruction can refer to the process by which the decoding network generates output data with the same dimension as the original input data based on the low-dimensional data.

[0067] Based on the stacked autoencoder part in the system embodiment, its interpretability guarantee mechanism is specifically implemented through the collaborative training of the encoding network and the decoding network.

[0068] The encoding network consists of multiple fully connected layers connected in sequence. Each layer performs a non-linear transformation on the input and outputs a feature representation with lower dimensionality. The output of the last encoding layer is the dimensionality-reduced data.

[0069] The decoding network is a mirror-symmetric structure of the encoding network, and its input is the aforementioned dimensionality-reduced data. Each layer in the decoding network performs the opposite operation to the encoding layer. Specifically, unpooling is used in the middle layers to recover the spatial structure information of the data, and deconvolution is used in the layers near the output to finely reconstruct the detailed features of the data. The final output of the decoding network is the reconstructed data, whose dimensions are exactly the same as the original multimodal input data.

[0070] During training, the loss function is defined as the difference (e.g., mean squared error) between the reconstructed data and the original multimodal data. The backpropagation algorithm continuously optimizes all parameters of the encoding and decoding networks, aiming to minimize this reconstruction error. This training process forces the encoding network to learn the most essential and crucial information from the original data and compress it into the dimensionality-reduced data. Only in this way can the decoding network effectively reconstruct the original data using this limited information. By observing whether the decoding network can successfully reconstruct the data and analyzing the contribution of each dimension of the dimensionality-reduced data to the reconstruction of different original features, we can qualitatively and quantitatively assess which key features are preserved. This ensures that the entire dimensionality reduction process is not a "black box," and its output has a clear physical or business meaning.

[0071] The beneficial effect of this embodiment is that by introducing a decoding network with reconstruction capabilities and jointly training it with the encoding network, interpretability constraints are added to the data dimensionality reduction process. This method ensures that the dimensionality-reduced data is not only small in size, but also faithfully retains the core features of the original data that are crucial to subsequent tasks, avoiding subsequent intelligent decision-making errors due to information loss or distortion, and enhancing the reliability and credibility of the entire data processing chain.

[0072] In one possible implementation, the decision execution module employs reinforcement learning algorithms, fuzzy logic algorithms, or rule base algorithms for local collaborative decision control.

[0073] Among these, reinforcement learning algorithms can refer to machine learning methods where an agent learns optimal behavioral strategies by obtaining reward signals through interaction with the environment. Fuzzy logic algorithms can refer to mathematical methods for handling uncertainty and approximate reasoning, transforming fuzzy inputs into explicit outputs through membership functions and fuzzy rules. Rule base algorithms can refer to methods for logical reasoning and decision-making based on pre-defined "condition-action" pairs (IF-THEN rules).

[0074] Based on the agent decision-making and execution module in the system embodiment, this module can use a variety of algorithms to make decisions, and the specific implementation methods are as follows.

[0075] Rule-based algorithms are a fundamental implementation approach. The decision execution module pre-stores a series of expert rules in the form of "IF-THEN". For example, "IF (real-time electricity price > set threshold) AND (battery state of charge > 50%) THEN (execute discharge strategy)". When the structured cognitive output from the fusion cognitive module matches the condition of a rule, the corresponding control action is triggered. This method is simple, reliable, and suitable for scenarios with clearly defined logic.

[0076] Fuzzy logic algorithms are used to process cognitive information with uncertainty. For example, the cognitive judgment of "grid voltage support demand" may be a fuzzy concept of "high," "medium," or "low." The fuzzy inference engine in the decision execution module transforms fuzzy cognitive inputs into explicit control outputs (such as reactive power reference values) based on the set membership functions and fuzzy rule base, enabling decisions to better adapt to operating states with unclear boundaries.

[0077] Reinforcement learning algorithms endow agents with the ability to learn and optimize through continuous interaction with the environment. The decision-making and execution module treats itself as an agent, taking the environmental state (i.e., the output of the fusion cognition module) as input, the executed control actions (such as adjusting output) as output, and converting superior coordination instructions or local goals (such as maximizing profit) into reward signals. Through continuous iteration of "state-action-reward," the agent eventually learns an optimal policy function, enabling it to make near-optimal decisions when faced with complex cognitive states it has never experienced before. In practical deployment, an offline cloud-based training and online inference on the agent can be adopted.

[0078] The beneficial effects of this embodiment are that the decision execution module supports multiple algorithmic paradigms, providing high flexibility and adaptability. The rule base ensures the determinism of basic behaviors; fuzzy logic enhances the ability to handle uncertainty; and reinforcement learning opens up a path for autonomous evolution and optimization. Multiple algorithms can be used individually or in combination, enabling the agent to select an appropriate decision-making mechanism based on specific application scenarios, computing resources, and performance requirements, balancing the real-time performance, accuracy, and intelligence of the decision.

[0079] In one possible implementation, the network controller is also used to implement access control policies, traffic monitoring, and intrusion detection.

[0080] Access control policies refer to a set of rules defined and enforced by the network controller, specifying which network entities (such as agents) can access which network resources under what conditions. Traffic monitoring refers to the process by which the network controller continuously collects, statistically analyzes, and processes communication data packets flowing through the data plane. Intrusion detection refers to the process by which the network controller automatically identifies potential malicious activities or security policy violations in the network by analyzing traffic and behavioral patterns.

[0081] Based on the software-defined network controller in the system embodiment, its security functions are specifically implemented as follows.

[0082] Access control policy: The network controller maintains an agent whitelist and role permission matrix. When a new agent device attempts to access the network, the controller first verifies its digital certificate or MAC address. Only authenticated agents are included in the global network view and assigned corresponding communication permissions based on their roles (such as "photovoltaic power generation unit" or "critical protection equipment"). For example, only agents acting as critical protection equipment are allowed to send the highest priority emergency shutdown command to the network controller.

[0083] Traffic monitoring: The controller continuously collects port traffic statistics and packet samples from various data plane devices. It runs traffic analysis applications to establish a normal communication traffic baseline model (such as the frequency and size of data uploaded periodically by various agents). Once an abnormal traffic pattern is detected, such as an agent sending a massive number of packets in a short period of time (potentially maliciously controlled to become a source of DoS attacks), or communication content not conforming to established protocol specifications, the controller will immediately flag the anomaly.

[0084] Intrusion Detection: Combining traffic monitoring results, the controller integrates both feature-based and anomaly-based intrusion detection engines. When malicious traffic matching a known attack signature database is detected, or when communication behavior deviates significantly from the baseline model, the intrusion detection engine determines that a security event has occurred. The controller then initiates response mechanisms, such as alerting the administrator and automatically issuing flow table rules to redirect traffic generated by the suspected compromised agent to an isolation zone for analysis, or directly blocking its communication connections with other critical devices in the network to prevent the attack from spreading.

[0085] The beneficial effect of this embodiment is that by integrating centralized security management functions into the network controller, unified and proactive defense against network security threats to the information and energy system's communication network is achieved. Access control eliminates unauthorized access at the source; traffic monitoring provides global situational awareness; and intrusion detection and automated response can quickly isolate threats, ensuring that the overall security and reliability of the entire collaborative communication network are not fundamentally affected when some nodes are attacked.

[0086] In one possible implementation, a network swarm optimization layer is set up between the fog computing layer and the cloud computing layer. The network swarm optimization layer uses swarm intelligence optimization algorithms to optimize network resource allocation.

[0087] Among them, the network swarm optimization layer can refer to a logical functional layer whose algorithms are distributed and run across the aforementioned software-defined network controller and multiple fog computing nodes, responsible for solving large-scale distributed optimization problems. Swarm intelligence optimization algorithms can refer to a class of metaheuristic optimization algorithms inspired by the cooperative behavior of biological groups (such as flocks of birds or ant colonies), where collective intelligence emerges through simple individual interactions, such as particle swarm optimization and ant colony optimization.

[0088] Based on the three-layer architecture of the system implementation, a logical network cluster optimization layer is further introduced between fog computing and cloud computing. This layer is not an independent physical device layer, but its functionality is implemented in the collaborative software of the software-defined network controller and fog computing nodes.

[0089] Its core task is to solve large-scale distributed optimization problems. For example, when the cloud issues a macro-level goal of "peak shaving and valley filling across the entire network," directly solving an optimization problem involving tens of thousands of variables would be computationally intensive. This is where the network swarm optimization layer comes into play.

[0090] The software-defined network controller utilizes its global network view to divide the entire system into multiple sub-regions that are highly coupled electrically and communicatively. Each sub-region is headed by a dominant fog computing node. The optimization problem is decomposed into local optimization problems within each sub-region.

[0091] Each "group leader" fog node employs a particle swarm optimization algorithm to solve the optimization model for its region. Each "particle" represents a set of possible power output schemes for the agent within the region, and its "position" is iteratively updated based on its own historical best and the best information of other particles in the swarm, ultimately searching for a locally optimal solution. During the iteration process, "group leaders" in adjacent regions exchange boundary coupling information (such as tie-line power) through a software-defined network to coordinate their solutions.

[0092] Through this distributed iteration based on swarm intelligence, all regions eventually converge to a solution close to the global optimum. The network swarm optimization layer dynamically adjusts the optimization boundaries and coordination strategies for each region to ensure efficient convergence of the algorithm. The optimization results are decomposed by each fog computing node and distributed to the agents within their respective regions for execution.

[0093] The beneficial effect of this embodiment is that by introducing a network swarm optimization layer and employing a swarm intelligence optimization algorithm, it provides an efficient and scalable solution for handling ultra-large-scale, distributed, and non-convex energy optimization problems. It avoids the "curse of dimensionality" problem of centralized computing, and through swarm collaboration and distributed iteration, it significantly reduces computational complexity and communication overhead while ensuring solution accuracy. This is a key technical aspect for achieving the global economic optimization goal of the system.

[0094] In one embodiment, the system hardware deployment architecture is as follows.

[0095] Terminal Layer (Self-managed): Deploys various embedded intelligent controllers, integrated into devices such as photovoltaic inverters, energy storage converters, smart switches, and smart meters. The hardware needs to have multi-channel data acquisition interfaces, necessary computing cores (such as ARM Cortex A series), and communication modules (4G / 5G, WiFi, LoRa, power line carrier).

[0096] Aggregation Layer (Interconnection): Fog computing gateways (using industrial-grade industrial control computers or high-end edge computing boxes) are deployed in locations such as power distribution rooms and base stations to aggregate regional data and run local optimization algorithms.

[0097] Network layer (cluster): Deploy commercial or custom SDN switches, routers, and wireless access points. The central data center deploys high-performance servers as SDN controllers.

[0098] Cloud: Deploying virtual machines or container clusters in public or private cloud data centers to run cloud platform management software, big data analytics engines, and AI training frameworks.

[0099] Furthermore, the software platform and algorithm implementation are described in detail below.

[0100] The core software for the intelligent agent employs a lightweight operating system (such as Linux RT) and uses C / C++ and Python to write the core control logic. It integrates a lightweight AI inference engine (such as TensorFlow Lite, ONNX Runtime).

[0101] Data Processing and AI Services: Implement DSAE and RBF networks using Python / PyTorch / TensorFlow. Utilize Apache Spark or Flink for streaming data processing.

[0102] SDN controller: Based on open source platforms (such as ONOS, OpenDaylight), secondary development is carried out to write northbound APIs and applications.

[0103] Cloud platform and simulation engine: Adopting a microservice architecture and developed using Java / Go. The simulation core can be based on MATLAB / Simulink, OPALRT, or a self-developed parallel simulation framework.

[0104] Cooperative control algorithms: Implement distributed optimization algorithms based on consensus algorithms and alternating direction multiplier method (ADMM), as well as cloud-based policy training algorithms based on deep reinforcement learning (DRL).

[0105] The specific workflow and simulation verification are as follows.

[0106] Physical entities are registered as virtual agents within the system. These agents continuously collect multimodal data, which is then fused and feature extracted locally or at fog nodes. In Scenario 1 (voltage stability): A sudden surge in photovoltaic output in a certain area leads to voltage exceeding limits. Upon detection, the local agent can autonomously activate energy storage to absorb reactive power or adjust the inverter's power factor (autonomous). If this fails, the fog computing node coordinates with nearby energy storage and controllable loads for collaborative voltage regulation (mutual).

[0107] In Scenario 2 (Economic Scheduling): The cloud generates the optimal scheduling plan for the next day based on weather forecasts and electricity price signals, and distributes it to each fog node and intelligent agent (cloud). The fog nodes make fine-tuning adjustments based on real-time conditions.

[0108] Simulation Testing: A virtual urban power distribution network was constructed, comprising 10,000 residential solar panels, 2,000 energy storage devices, and 500 charging piles. A fault scenario of "large-scale solar panel detachment under extreme weather conditions" was set up. Traditional control strategies and the platform's agent-cooperative strategy were tested separately, and the effectiveness of the platform was verified by comparing indicators such as system recovery time, load loss, and voltage recovery. A prototype system has been built using tools such as CloudSim, OMNeT++, and Python. Simulation tests were conducted on core algorithms (such as DSAE dimensionality reduction and agent-cooperative algorithms), and the results verified their effectiveness and performance advantages.

[0109] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that the embodiments in this specification are not limited to the described order of actions, because according to the embodiments in this specification, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in this specification are all preferred embodiments, and the actions and modules involved are not necessarily essential to the embodiments in this specification.

[0110] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0111] The preferred embodiments disclosed above are merely illustrative of this specification. The optional embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the embodiments described herein. These embodiments are selected and specifically described in this specification to better explain the principles and practical applications of the embodiments, thereby enabling those skilled in the art to better understand and utilize this specification. This specification is limited only by the claims and their full scope and equivalents.

Claims

1. A multi-agent cooperative simulation control system for an information energy system, characterized in that, include: Multiple intelligent agents are deployed in the schedulable unit of the information energy system to collect multimodal raw data from the schedulable unit and perform local collaborative decision-making and control. The data processing module, connected to the intelligent agent, is used to perform interpretable dimensionality reduction processing on the multimodal raw data to obtain dimensionality-reduced data, and to perform feature dimensionality augmentation processing on the dimensionality-reduced data to obtain dimensionality-enhanced feature data. A software-defined network connects the intelligent agent and the data processing module, and is used to provide dynamic resource scheduling and secure communication for the interaction between the intelligent agents; The collaborative management and control simulation platform connects the intelligent agent, the data processing module, and the software-defined network. It is used to integrate edge computing nodes, fog computing nodes, and cloud computing nodes to form a hierarchical control system, and to perform closed-loop simulation of the system containing the intelligent agent, the data processing module, and the software-defined network.

2. The multi-agent cooperative simulation control system for information energy systems according to claim 1, characterized in that, The data processing module includes a stacked autoencoder and a radial basis function network. The stacked autoencoder is used to reduce the dimensionality of the original multimodal data, and the radial basis function network is used to perform nonlinear mapping on the reduced dimensionality data to obtain the increased dimensionality feature data.

3. The multi-agent cooperative simulation control system for information energy systems according to claim 1, characterized in that, The intelligent agent includes a perception module, a fusion cognition module, and a decision execution module. The perception module is used to collect electrical quantities, physical quantities, state quantities, and environmental quantities. The fusion cognition module is used to extract and fuse features from the collected data. The decision execution module is used to make local decisions and perform collaborative control based on the fused cognition.

4. The multi-agent cooperative simulation control system for information energy systems according to claim 1, characterized in that, The software-defined network adopts a control and forwarding separation architecture, including a control plane and a data plane. The control plane is equipped with a network controller to realize a global network view and resource scheduling. The data plane consists of distributed network devices and the communication interface of the agent.

5. The multi-agent cooperative simulation control system for information energy systems according to claim 1, characterized in that, The collaborative management and control simulation platform includes an edge computing layer, a fog computing layer, and a cloud computing layer. The edge computing layer is used to realize the rapid local control of the intelligent agent, the fog computing layer is used to coordinate multiple intelligent agents to perform local optimization, and the cloud computing layer is used for data analysis and model training.

6. The multi-agent cooperative simulation control system for an information energy system according to claim 5, characterized in that, The collaborative management and control simulation platform has a built-in simulation engine, which is used to simulate network topology containing energy nodes and supports setting up operating scenarios to test different control strategies.

7. The multi-agent cooperative simulation control system for information energy systems according to claim 2, characterized in that, The stacked autoencoder includes an encoding network and a decoding network. The decoding network is used to reconstruct the dimensionality-reduced data and retain interpretable features by comparing the reconstructed data with the original multimodal data.

8. The multi-agent cooperative simulation control system for information energy systems according to claim 3, characterized in that, The decision execution module employs reinforcement learning algorithms, fuzzy logic algorithms, or rule base algorithms for local collaborative decision control.

9. The multi-agent cooperative simulation control system for an information energy system according to claim 4, characterized in that, The network controller is also used to implement access control policies, traffic monitoring, and intrusion detection.

10. The multi-agent cooperative simulation control system for an information energy system according to claim 5, characterized in that, A network swarm optimization layer is also provided between the fog computing layer and the cloud computing layer. The network swarm optimization layer uses a swarm intelligence optimization algorithm to optimize network resource allocation.

Images