Optimization control method of power system

Abstract

The invention discloses an optimization control method for a power system, and relates to the technical field of power control, and the method comprises the steps: constructing a multi-energy state space, and building a layered optimization architecture based on the multi-energy state space; the decision-making layer generates a multi-energy balance scheduling instruction through a multi-target collaborative optimization algorithm to realize multi-target collaborative optimization of the electric heating gas system; the execution layer comprises a plurality of energy subsystem control modules, and collaborative strategy optimization is carried out through a distributed reinforcement learning algorithm; positioning a disturbance source of topology mutation through a causal reasoning algorithm, and generating a reconstruction instruction; training a control strategy of each energy subsystem by adopting a distributed learning framework, and executing a distributed cooperative control action according to the reconstruction instruction; the network connectivity is judged through the communication state of the energy subsystem, and if the communication is normal, collaborative decision making is carried out through a distributed learning framework; and if the communication is interrupted, switching to an edge node offline autonomous mode.

Description

Technical Field

[0001] This invention relates to the field of power control technology, and in particular to an optimized control method for power systems. Background Technology

[0002] Existing power system optimization control methods suffer from fundamental flaws in multi-energy coordination, primarily manifested in unbalanced scheduling of multiple energy subsystems and low optimization efficiency. Due to the complex and variable operating environment of electro-thermal-gas coupled systems and the dynamic uncertainty of network topology, existing technologies struggle to establish a unified state representation system, resulting in insufficient system response to sudden disturbances. In wide-area distributed network environments with multi-energy coupling, systems must simultaneously address the challenges of fluctuating physical conditions and real-time data acquisition; existing methods lack effective hierarchical coordination mechanisms and cannot adapt to the optimization needs under complex operating conditions. System operators seek a balance between economy and reliability, end-users prioritize energy supply stability, and equipment manufacturers focus on cost control and maintainability. Existing technologies struggle to coordinate these differentiated needs, often sacrificing performance in other areas to meet the requirements of one. This ultimately manifests as systematic deviations in several key indicators: unstable convergence of energy dispatch commands, difficulty in guaranteeing communication transmission quality, insufficient accuracy in fault location, poor system response timeliness, and overall energy efficiency falling short of expectations. Summary of the Invention

[0003] This application provides an optimized control method for power systems, which solves the problems of unbalanced scheduling of multiple energy subsystems, lag in response to dynamic topology changes, and insufficient robustness of the system under communication interruption in the prior art. It achieves the technical effects of improving the convergence speed of scheduling commands, shortening the system recovery time to the minute level under extreme events, and improving the comprehensive energy efficiency of multiple energy sources.

[0004] This application provides an optimal control method for a power system, including:

[0005] S1: Construct a multi-energy state space to collect real-time operational status data of the power system, heating system, and natural gas system; establish a hierarchical optimization architecture based on the multi-energy state space, including a high-level multi-energy collaborative decision-making layer and a low-level cross-domain distributed execution layer;

[0006] S2: The decision-making layer generates multi-energy balanced scheduling instructions through a multi-objective collaborative optimization algorithm to achieve multi-objective collaborative optimization of the electric, heating, and gas systems; the execution layer contains multiple energy subsystem control modules and performs collaborative strategy optimization through a distributed reinforcement learning algorithm.

[0007] S3: Based on the multi-energy state space, the disturbance source of topological change is located through causal reasoning algorithm and reconfiguration command is generated; the control strategy of each energy subsystem is trained using a distributed learning framework and distributed cooperative control actions are executed according to the reconfiguration command.

[0008] S4: Based on the reconfiguration instructions and control strategies, initiate a multi-level response mechanism; determine network connectivity by checking the communication status of the energy subsystem; if communication is normal, make collaborative decisions through a distributed learning framework; if communication is interrupted, switch to the offline autonomous mode of the edge node, call the scenario strategy library to execute distributed control actions, and restore system functionality.

[0009] Furthermore, the multi-energy state space includes: mapping the operational state data of the power system, heating system, and natural gas system into a unified state vector through standardization processing.

[0010] ,

[0011] in, To unify the state vector, For the power grid state sub-vector, This is the state sub-vector of the heating network. Let the gas network state sub-vector be... This is the topological connection state sub-vector.

[0012] Furthermore, the multi-objective cooperative optimization algorithm models the electrical, thermal, and gas subsystems as independent intelligent agents, and generates multi-energy equilibrium scheduling instructions through Nash equilibrium solution:

[0013] ,

[0014] in, Let be the optimal policy of agent i in the (k+1)th iteration. Let be the utility function of agent i. Let i be the policy to be evaluated. For the historical policies of all other agents at the k-th iteration, To enable the function The variable that achieves the maximum value The value;

[0015] When the policy change is less than the threshold, output the Nash equilibrium policy combination:

[0016] ,

[0017] in, For Nash equilibrium strategy combinations, The optimal action strategy for the power grid subsystem. The optimal action strategy for the heating network subsystem. The optimal action strategy for the gas network subsystem;

[0018] Transform the Nash equilibrium strategy combination into a multi-energy equilibrium scheduling instruction;

[0019] The threshold is determined using historical data and has a value range of [0.001, 0.01].

[0020] Furthermore, the distributed reinforcement learning algorithm is an improved SAC algorithm, which embeds multi-energy physical constraint terms into the reward function:

[0021] ,

[0022] in, For the total reward function, As a basic economic incentive item, This is a penalty term for power grid frequency deviation. For the temperature constraint term of the heating network, This is a safety threshold constraint term for gas network pressure.

[0023] Furthermore, the causal reasoning algorithm employs the PC algorithm to construct a catastrophic causal graph based on historical event data, predicts the fault propagation path of topological mutations based on the causal graph, and generates a pre-intervention instruction set that includes fault isolation, energy transfer, and redundancy activation.

[0024] Furthermore, the distributed learning framework is a federated learning architecture. Each energy subsystem trains policy model parameters locally and periodically uploads the encrypted policy model parameters to the parameter aggregation server. The parameter aggregation server aggregates parameters from multiple subsystems using a federated averaging algorithm to generate global model parameters. The global model parameters are processed by a differential privacy mechanism and then distributed to each energy subsystem for local model updates.

[0025] Furthermore, the multi-level response mechanism includes a three-level response: the first-level response performs fault isolation based on causal reasoning within 0-5 minutes; the second-level response restores critical loads through offline autonomous recovery of edge nodes within 5-30 minutes; and the third-level response performs rolling optimization through federated learning after communication is restored.

[0026] Furthermore, the communication status judgment includes real-time monitoring of network latency and packet loss rate. When the latency exceeds a preset threshold and the packet loss rate exceeds a critical value, it is determined that the communication is interrupted. Then, the edge node automatically switches to offline autonomous mode and calls the pre-stored scenario policy library to execute control actions. The scenario policy library contains policies pre-trained through digital twin simulation based on historical event data to ensure that system functions are quickly restored when communication is interrupted.

[0027] Furthermore, the scenario policy library contains pre-trained strategies for extreme scenarios, generated through simulation in a digital twin environment. This includes: constructing a digital twin environment, integrating historical extreme event data and a physical model of a multi-energy system, simulating the dynamic behavior of an electrothermal-gas coupling system, and introducing sensor data to generate extreme scenario sequences; training the strategies in the digital twin environment using an improved SAC reinforcement learning algorithm, verifying the effectiveness of the training results, and after successful verification, compressing and storing them in the local policy library of the edge node, and classifying and labeling them as specific extreme scenarios.

[0028] Furthermore, the verification of the training results includes: verifying the training strategy; when the strategy meets the predefined performance indicators, compressing and storing it in the local strategy library of the edge node; when communication is interrupted, the edge node calls the matching pre-trained strategy from the local strategy library according to the current state to execute control actions; the performance indicators include a recovery time reduction rate of not less than 50% and a critical load guarantee rate of not less than 95%.

[0029] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0030] By employing a multi-objective collaborative optimization algorithm to generate balanced scheduling instructions, combining distributed reinforcement learning for policy optimization, utilizing causal reasoning to locate disturbance sources and generate reconstruction instructions, implementing distributed training and privacy protection through a federated learning framework, and designing a multi-level response and communication adaptive mechanism, the technical effects of achieving balanced multi-energy scheduling, minute-level response to topology changes, enhanced system resilience, and improved overall energy efficiency are realized. Attached Figure Description

[0031] Figure 1 This is a flowchart of an optimized control method for a power system according to an embodiment of the present invention. Detailed Implementation

[0032] To facilitate understanding of the present invention, a more complete description of this application will be given below with reference to the accompanying drawings, which illustrate preferred embodiments of the invention. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to enable a more thorough and complete understanding of the disclosure of the present invention.

[0033] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains; the terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to limit the invention; the term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0034] Example 1: As Figure 1 As shown, an optimization control method for a power system is proposed.

[0035] S1: Construct a multi-energy state space to collect real-time operational status data of the power system, heating system, and natural gas system; establish a hierarchical optimization architecture based on the multi-energy state space, including a high-level multi-energy collaborative decision-making layer and a low-level cross-domain distributed execution layer;

[0036] Specifically, the system collects real-time data on power systems such as node load, renewable energy output, energy storage device status of charge, grid frequency, and voltage deviation; thermal systems such as pipeline water temperature, flow rate, heat exchange station valve opening, heat load demand, and heat exchange efficiency; and natural gas systems such as pipeline pressure, gas source flow rate, gas storage tank inventory levels, compressor status, and user gas demand. The multimodal data is normalized and smoothed using a sliding time window to reduce noise interference and ensure data stability.

[0037] The multi-energy state space includes mapping the operational state data of the power system, heating system, and natural gas system into a unified state vector through standardization processing.

[0038] ,

[0039] in, To unify the state vector, This is a sub-vector representing the grid state, containing node load sequences, renewable energy output forecasts, and energy storage state of charge, used to reflect the real-time operating status of the grid. This is a sub-vector representing the state of the heating network, containing information such as pipe water temperature distribution, heat load demand, and valve opening, used to reflect the energy flow state of the heating network. This is a sub-vector representing the state of the gas network, containing dimensions such as pipeline pressure gradient, gas supply, and dynamic inventory in gas storage tanks, used to reflect the supply and demand balance of the gas network. The topology connection state sub-vector dynamically encodes the physical connection relationships of the electricity, heat, and gas networks: within the power network, an adjacency matrix is ​​used to encode the power network switching state (0 / 1 represents on / off) and microgrid grid-connected / off-grid flags; within the heat network, an incidence matrix is ​​used to describe the valve opening degree (continuous values ​​0~1) and pipeline connectivity; within the natural gas network, a tagging mechanism is used to record the operating mode of the gas network compressor and the state of the gas storage tank node; based on sensor data refreshed every 5 seconds, a topology fingerprint is generated using a graph hashing algorithm.

[0040] S2: The decision-making layer generates multi-energy balanced scheduling instructions through a multi-objective collaborative optimization algorithm to achieve multi-objective collaborative optimization of the electric, heating, and gas systems; the execution layer contains multiple energy subsystem control modules and performs collaborative strategy optimization through a distributed reinforcement learning algorithm.

[0041] Specifically, based on a multi-energy state space, a hierarchical optimization architecture is constructed. The high-level multi-energy collaborative decision-making layer models the electricity, heat, and gas subsystems as three independent intelligent agents, each pursuing its own optimization objective: the power grid agent aims to minimize operating costs and frequency deviation, the heating network agent aims to maximize heating efficiency and temperature stability, and the gas network agent aims to maintain pressure safety margin and supply reliability. Multi-energy equilibrium scheduling instructions are generated through Nash equilibrium solution.

[0042] ,

[0043] in, Let be the optimal policy of agent i in the (k+1)th iteration. Let be the utility function of agent i. Let i be the policy to be evaluated. For the historical policies of all other agents at the k-th iteration, To enable the function The variable that achieves the maximum value The value;

[0044] In each iteration, agent i follows the historical strategies of other agents. Optimize their own strategies Maximize utility function The utility function covers economic, safety, and stability objectives: the utility function of the power grid intelligent agent includes electricity price cost and frequency deviation penalty; the utility function of the heating network intelligent agent includes heating efficiency and temperature fluctuation penalty; and the utility function of the gas network intelligent agent includes pressure over-limit risk and gas supply interruption cost.

[0045] When the policy change is less than the threshold, output the Nash equilibrium policy combination:

[0046] ,

[0047] in, For Nash equilibrium strategy combinations, The optimal action strategy for the power grid subsystem. The optimal action strategy for the heating network subsystem. The optimal action strategy for the gas network subsystem is determined by historical data, with a value range of [0.001, 0.01].

[0048] The Nash equilibrium strategy is combined and transformed into multi-energy equilibrium dispatch instructions: the power grid strategy is parsed into generator output setpoints, load switching instructions, and energy storage charging and discharging power; the heating network strategy is transformed into heat exchange station valve openings, pump frequencies, and heat source output adjustments; and the gas network strategy is mapped into compressor speed, gas storage tank output flow, and pipeline pressure regulating valve openings. The parsed control parameters are encapsulated into standardized instruction packages, including metadata such as timestamps, target values, execution priorities, and validity periods.

[0049] The lower-level cross-domain distributed execution layer comprises control modules for three energy subsystems: the power grid, heating network, and gas network. Each subsystem control module acts as an independent execution unit, receiving balance scheduling instructions and local state data from the decision-making layer. Each module optimizes its strategy using an improved SAC algorithm, expanding its state space to a joint input. ,in, This is a low-level local state. The high-level dispatch commands are used for the action space, which is continuously controlled by the output of each subsystem, such as the load migration amount of the grid actuator, the energy storage charging and discharging commands, the valve opening of the heating network actuator, and the compressor frequency of the gas network actuator.

[0050] Embed multi-energy physics constraints in the reward function:

[0051] ,

[0052] in, For the total reward function, As a basic economic incentive item, This is a penalty term for power grid frequency deviation. For the temperature constraint term of the heating network, This is a safety threshold constraint term for gas network pressure.

[0053] The overall economic objectives reflecting system operation are mainly calculated based on energy procurement costs, equipment operating losses, and energy efficiency indicators.

[0054] ,

[0055] in, The costs of purchasing electricity from the power grid and the costs of grid losses, For the cost of heating network supply, For the cost of gas supply to the gas network, For comprehensive energy efficiency indicators of multiple energy sources, such as electrothermal conversion efficiency and gas-fired power generation efficiency, weighting is used. Adjust its importance, The value range is [0.5, 2.0], determined by historical data.

[0056] To penalize deviations of the grid frequency from its rated value (e.g., 50Hz) and ensure power system stability, the calculation method is as follows:

[0057] ,

[0058] in, The real-time frequency deviation value is acquired through a synchronous phasor measurement unit; The penalty coefficient is determined based on the system's inertia constant and ranges from [0.1 to 1.0]. The larger the deviation, the more severe the penalty.

[0059] To ensure the water temperature in the heating network pipelines remains within a safe range and to prevent equipment scaling or interruption of heat load supply, the calculation method is as follows:

[0060] ,

[0061] in, For real-time pipeline water temperature, For the safe lower limit temperature, This is the temperature penalty coefficient, which is related to the thermal inertia of the heating network, and its value ranges from [0.05, 0.5].

[0062] Penalty measures are applied to gas pipeline pressure deviating from the safe range to prevent gas supply interruptions or pipeline damage. The calculation method is as follows:

[0063] ,

[0064] in, For real-time pipeline pressure, To the lower limit of safety pressure, This is the pressure penalty coefficient, which is related to the flow characteristics of the gas network, and its value ranges from [0.2 to 1.0].

[0065] The reward function guides the policy network updates: the agent explores the action space through trial and error, and the reward value after each action is used to calculate the advantage function, driving the policy to optimize towards higher rewards. Constraints, through a penalty mechanism, enable the policy to automatically avoid dangerous actions.

[0066] By embedding multi-energy physical constraints into the reward function, the execution layer achieves precise coordination within a layered architecture. For example, during the winter evening rush hour, the grid load surges, heating network demand increases, and gas network pressure is constrained. The decision layer generates a balanced dispatch command, requiring peak shaving for the grid, heat network insulation, and gas network pressure maintenance. The execution layer's grid agent is based on... and Prioritize the use of energy storage discharge to avoid frequency deviation; the intelligent heating network system is based on Adjusting valve opening to maintain water temperature, while simultaneously responding to grid commands to reduce electric heating load; Gas grid intelligent agent: based on Start the gas tank to compensate for the pressure and avoid over-reliance on the compressor.

[0067] The technical solutions described in the embodiments of this application have at least the following technical effects or advantages:

[0068] This application constructs a multi-energy state space and establishes a hierarchical optimization architecture based on it. It generates multi-energy balanced scheduling instructions through a multi-objective collaborative optimization algorithm and combines an improved SAC distributed reinforcement learning algorithm for collaborative strategy optimization. This achieves multi-objective collaborative optimization of the electric heating gas system, effectively improving the system's economy, frequency stability, temperature control accuracy, and pressure safety. Furthermore, the precise collaboration under the hierarchical architecture ensures the balanced scheduling of multiple energy sources and the stable operation of the system.

[0069] Example 2: While Example 1 achieved multi-objective collaborative optimization of the electric heating gas system, it still suffers from insufficient responsiveness to dynamic changes. This example further supplements Example 1.

[0070] S3: Based on the multi-energy state space, the disturbance source of topological change is located through causal reasoning algorithm and reconfiguration command is generated; the control strategy of each energy subsystem is trained using a distributed learning framework and distributed cooperative control actions are executed according to the reconfiguration command.

[0071] The causal reasoning algorithm adopts the PC algorithm, constructs a catastrophic causal graph based on historical event data, predicts the fault propagation path of topological mutation based on the causal graph, and generates a pre-intervention instruction set including fault isolation, energy transfer and redundancy activation.

[0072] Specifically, the PC algorithm is used to analyze the causal relationship between historical topology change events (such as microgrid grid connection / off-grid, line switch operation) and changes in the state of multiple energy sources. The algorithm uses statistical tests to determine the conditional independence between variables and generates a causal graph in the form of a directed acyclic graph. The nodes in the causal graph encompass the state variables of multiple energy sources, and the edges represent causal relationships. When a topology change signal is detected, the causal graph quickly locates the disturbance source. For example, if a microgrid off-grid event causes a grid frequency deviation, the PC algorithm can trace the root cause back to the topology connection state.

[0073] Based on the cause-effect graph, reconfiguration instructions are output, including: fault isolation instructions that automatically disconnect abnormal lines or nodes to prevent fault propagation; energy transfer instructions that transfer loads or energy supplies to safe paths; and redundancy activation instructions that activate energy storage devices or backup heat sources in the heating network to compensate for energy shortages. The instructions are encapsulated in a standardized format, including timestamps, target values, and execution priorities, ensuring seamless integration with the distributed execution layer.

[0074] The distributed learning framework is a federated learning architecture. Each energy subsystem trains policy model parameters locally and periodically uploads encrypted policy model parameters to a parameter aggregation server. The parameter aggregation server aggregates parameters from multiple subsystems using a federated averaging algorithm to generate global model parameters. The global model parameters are processed through a differential privacy mechanism and then distributed to each energy subsystem for local model updates.

[0075] Specifically, a federated learning architecture is adopted as the distributed learning framework to achieve collaborative training and optimization of control strategies for various energy subsystems while protecting data privacy. Each energy subsystem acts as a federated learning client, locally deploying an improved SAC policy model. Training is conducted every 5 minutes using the latest one-hour data from a sliding window, with a maximum of 100 training iterations, or early termination when the loss function changes by less than 0.001, to balance real-time performance and accuracy. Each subsystem periodically encrypts its local model parameters and uploads them to the parameter aggregation server. The AES-256 symmetric encryption algorithm is used to encrypt the model parameters, and the key is dynamically distributed through a public key infrastructure. Each subsystem generates a session key locally, encrypts it using the server's public key, and transmits it to ensure key security. The server uses a federated averaging algorithm to aggregate parameters and generate global model parameters.

[0076] ,

[0077] in, These are global model parameters. For local parameters, The number of subsystems is specified. Laplacian noise is added to the global parameters, with a privacy budget ε set to 0.1. The noise scale is inversely proportional to the parameter sensitivity. The noise-added global parameters are then distributed to each subsystem for local model incremental updates.

[0078] The reconfiguration command serves as the input condition for the strategy model, and each subsystem generates distributed control actions: the power grid subsystem adjusts the energy storage charging and discharging plan according to the energy transfer command, prioritizing critical loads; the heating network subsystem adjusts valve openings and switches to backup heat sources based on the redundancy activation command; and the gas network subsystem executes pressure safety commands, controlling compressor speed and gas tank output. Action conflicts are resolved through a distributed consensus protocol to ensure global consistency.

[0079] S4: Based on the reconfiguration instructions and control strategies, initiate a multi-level response mechanism; determine network connectivity by checking the communication status of the energy subsystem; if communication is normal, make collaborative decisions through a distributed learning framework; if communication is interrupted, switch to the offline autonomous mode of the edge node, call the scenario strategy library to execute distributed control actions, and restore system functionality.

[0080] The multi-level response mechanism includes a three-level response: the first-level response performs fault isolation based on causal reasoning within 0-5 minutes; the second-level response restores critical loads through offline autonomous recovery of edge nodes within 5-30 minutes; and the third-level response performs rolling optimization through federated learning after communication is restored.

[0081] Specifically, when the topology sensing layer detects a sudden event (such as a line fault or microgrid disconnection), it immediately initiates a Level 1 response. The causal reasoning module analyzes the event data in real time, locates the disturbance source, and generates a reconstruction command. The command is encapsulated in a standardized format and sent to the distributed execution layer to ensure initial isolation is completed within 5 minutes. After the Level 1 response, if the system does not fully recover or communication interruption causes collaborative decision-making to fail, a Level 2 response is initiated. Edge nodes automatically switch to offline autonomous mode, no longer relying on central communication; they call the locally pre-stored scenario policy library, match the optimal policy according to the current topology state, and execute distributed control actions. After communication is restored, a Level 3 response is triggered, and the system switches back to online mode; the federated learning framework is reactivated, and each energy subsystem uploads its locally trained parameters; the parameter aggregation server fuses the global model using a federated averaging algorithm and adds differential privacy noise to protect data security; the global model is sent to edge nodes for incremental updates to local policies, and the policies are dynamically adjusted through a reinforcement learning reward function to improve long-term performance.

[0082] The communication status judgment includes real-time monitoring of network latency and packet loss rate. When the latency exceeds a preset threshold and the packet loss rate exceeds a critical value, it is determined that the communication is interrupted. Then the edge node automatically switches to offline autonomous mode and calls the pre-stored scene policy library to execute control actions.

[0083] Specifically, the preset threshold for network latency is 100 milliseconds; anything exceeding this is considered communication latency. The critical value for packet loss rate is 5%; anything exceeding this is considered a communication interruption. Network metrics are collected every 10 seconds, and a sliding window averaging filter is used to reduce the impact of fluctuations. If both latency and packet loss rate exceed the limits, the system immediately triggers a communication interruption flag, and the edge node automatically switches to offline autonomous mode, calling the pre-stored scenario policy library to execute control actions. When the monitored metrics are below the threshold for three consecutive times, communication is considered restored, and a three-level response is initiated.

[0084] The technical solutions described in the embodiments of this application have at least the following technical effects or advantages:

[0085] This application introduces a causal reasoning algorithm to locate the disturbance source of topological mutations in real time and generates reconstruction instructions that include fault isolation, energy transfer, and redundancy activation. Combined with a federated learning architecture, it realizes distributed training and privacy protection of policies for each energy subsystem. At the same time, it designs a multi-level response mechanism and realizes automatic switching to offline autonomous mode under communication interruption based on real-time monitoring of network latency and packet loss rate. It calls the pre-stored digital twin simulation scenario policy library to execute control actions, thereby significantly improving the system's response speed and accuracy to dynamic topology changes. It solves the problems of unbalanced multi-energy scheduling and insufficient robustness under communication interruption, and achieves the technical effects of topology mutation recovery in minutes, multi-energy collaborative efficiency optimization, enhanced system resilience, and improved overall energy efficiency.

[0086] Example 3: Example 2 achieved dynamic topology response and distributed collaborative optimization of multi-energy systems, but still suffers from insufficient resilience under extreme events. This example further supplements Example 2.

[0087] The scenario policy library contains pre-trained strategies for extreme scenarios, generated through simulation in a digital twin environment. This includes: constructing a digital twin environment, integrating historical extreme event data and a physical model of a multi-energy system, simulating the dynamic behavior of an electrothermal-gas coupling system, and introducing sensor data to generate extreme scenario sequences; training the strategies in the digital twin environment using an improved SAC reinforcement learning algorithm, verifying the effectiveness of the training results, and after successful verification, compressing and storing them in the local policy library of the edge nodes, and classifying and labeling them as specific extreme scenarios.

[0088] Specifically, the digital twin environment is a high-fidelity virtual simulation platform used to simulate the dynamic behavior of multi-energy coupled systems involving electricity, heat, and gas. A historical extreme event database integrates extreme event data, such as earthquake magnitude, flood inundation range, typhoon path, and equipment failure records. Data sources include meteorological departments, power system fault databases, and user reports. The database is categorized and stored according to event type, intensity, and impact range, and aligned using timestamps to ensure data temporal consistency. The multi-energy system physical model includes a power grid model, a heating network model, a gas network model, and coupling interfaces. The power grid model simulates the dynamic response of generators, transmission lines, and loads; the heating network model simulates changes in pipeline water temperature, flow rate, and valve opening; and the gas network model simulates the dynamics of pipeline pressure, compressor speed, and gas storage tank inventory. The coupling interfaces define the interaction rules for the electricity, heat, and gas systems. Real-time sensor data from the physical system is used to calibrate the twin model parameters and reduce model drift. Data is sampled once per second, and noise is removed using filtering algorithms. Based on historical extreme event data and real-time sensor input, extreme scenario sequences are generated. For example, when simulating typhoon events, the module combines wind speed and precipitation forecasts to dynamically generate a cascading fault sequence of power grid pole collapse, gas grid pressure collapse, and heating grid interruption. The sequence time step is set to 5 seconds to capture the rapid evolution process.

[0089] The improved SAC reinforcement learning algorithm is used for policy training, and the state vector is expanded as follows:

[0090] ,

[0091] in, Identify extreme events to enable the strategy to distinguish between different extreme scenarios.

[0092] In addition to the reward function, a resilience reward term is added to encourage rapid recovery and minimizing impact:

[0093] ,

[0094] in, As a resilience reward item, For system recovery time, For critical load guarantee rate, Let be the weighting coefficient, with a value range of [0.5, 1.0]; the total reward function becomes: .

[0095] In each extreme scenario, the policy network is state-based. The input is the control action, and the output is the policy network. Through trial and error, the policy network is gradually optimized, aiming to maximize the cumulative reward while satisfying multiple energy constraints. The maximum number of training iterations is 1000, or it stops when the reward function converges (the change is less than 0.001). A parallel training architecture is adopted to handle multiple extreme scenarios simultaneously, improving training efficiency.

[0096] The verification of the training results includes: verifying the training strategy; when the strategy meets the predefined performance indicators, compressing it and storing it in the local strategy library of the edge node; when communication is interrupted, the edge node calls the matching pre-trained strategy from the local strategy library according to the current state to execute the control action.

[0097] Specifically, after training, the strategy is validated in a twin environment: the effectiveness of the strategy is checked to see if it can isolate faults and restore critical loads within a specified time (e.g., <10 minutes). Abnormal conditions such as sensor data loss and communication interruption are simulated to evaluate the robustness of the strategy. A recovery time reduction rate of ≥50% and a critical load assurance rate of ≥95% are considered passing thresholds.

[0098] Validated policies are compressed into lightweight parameter packages and stored in the local policy library of the edge nodes. Each policy package is labeled with metadata, including: scene label, applicable topology conditions, and policy priority.

[0099] The scenario policy library is deployed on each edge node, with a preset storage capacity of 100 policy packages. When an extreme event occurs and communication is interrupted, the edge node matches the current state with the policy label through its local causal inference module and automatically invokes the optimal pre-trained policy to execute control actions. After communication is restored, the federated learning platform aggregates new data, retrains the policy, and incrementally updates it to the edge nodes.

[0100] The technical solutions described in the embodiments of this application have at least the following technical effects or advantages:

[0101] This application constructs a high-fidelity digital twin environment, integrating historical extreme event data and a physical model of a multi-energy system to simulate the dynamic behavior of an electrothermal-gas coupled system, and introduces real-time sensor data to generate extreme scenario sequences. An improved SAC reinforcement learning algorithm is used for policy training, incorporating extreme event identifiers into the extended state vector and adding a resilience reward term to optimize recovery performance. After training, rigorous verification ensures the policy meets performance indicators of a recovery time reduction rate of no less than 50% and a critical load assurance rate of no less than 95%. The policy is then compressed into a lightweight parameter package and stored in the local policy library of the edge nodes. When communication is interrupted, the edge nodes automatically match the current state with the policy label, call the optimal pre-trained policy to execute distributed control actions, and incrementally update the policy through a federated learning platform after communication is restored. This significantly improves the system's resilience under extreme events, solves the problems of response lag and insufficient recovery capability of multi-energy systems in extreme scenarios, and achieves the technical effects of minute-level rapid recovery, reliable assurance of critical loads, and comprehensive energy efficiency optimization.

[0102] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. For those skilled in the art, the present invention can have various modifications and variations. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method of optimal control of an electric power system, characterized by, Comprise: S1: Construct a multi-energy state space, real-time collection of power system, heat system and natural gas system operation state data; Based on the multi-energy state space, a hierarchical optimization architecture is established, including a high-level multi-energy collaborative decision layer and a low-level cross-domain distributed execution layer; S2: The decision layer generates multi-energy balanced scheduling instructions through a multi-objective collaborative optimization algorithm, achieving multi-objective collaborative optimization of the electric-thermal-gas system; the execution layer includes multiple energy subsystem control modules, which optimize collaborative strategies through a distributed reinforcement learning algorithm; S3: Based on the multi-energy state space, the disturbance source of topological mutation is located through a causal reasoning algorithm, and reconstruction instructions are generated; the control strategies of each energy subsystem are trained using a distributed learning framework, and distributed collaborative control actions are executed according to the reconstruction instructions; S4: According to the reconstruction instructions and control strategies, a multi-level response mechanism is started; the network connectivity is determined through the communication state of the energy subsystem, if the communication is normal, the distributed learning framework is used for collaborative decision-making; If the communication is interrupted, switch to the edge node offline autonomous mode, call the scene strategy library to execute distributed control actions to restore system functions.

2. The method of optimal control of an electric power system according to claim 1, characterized in that, The multi-energy state space includes: the operation state data of the power system, the heat system and the natural gas system are mapped to a unified state vector through standardization processing: ， wherein, is a unified state vector, is a power grid state sub-vector, is a heating grid state sub-vector, is a gas grid state sub-vector, is a topological connection state sub-vector.

3. The method of optimal control of an electric power system according to claim 1, wherein The multi-objective collaborative optimization algorithm models the electric, thermal and gas subsystems as independent agents, and generates multi-energy balanced scheduling instructions through Nash equilibrium solution: ， wherein, is the optimal policy of agent i at iteration k + 1, is the utility function of agent i, is the policy to be evaluated for agent i, is the history policy of all other agents at iteration k, is the variable that maximizes the function the value of the variable​ When the strategy change is less than the threshold value, output the Nash equilibrium strategy combination: ， wherein, is the Nash equilibrium strategy combination, is the optimal action strategy for the power grid subsystem, is the optimal action strategy for the heating grid subsystem, is the optimal action strategy for the gas grid subsystem; Convert the Nash equilibrium strategy combination into multi-energy balanced scheduling instructions; The threshold value is calibrated through historical data, with a value range of [0.001, 0.01].

4. The method of optimal control of an electric power system according to claim 1, wherein, The distributed reinforcement learning algorithm is an improved SAC algorithm, which embeds multi-energy physical constraint terms in the reward function: ， wherein, is a total reward function, is a base economic reward term, is a grid frequency deviation penalty term, is a heat grid temperature constraint term, is a gas grid pressure safety threshold constraint term.

5. The method of optimal control of an electric power system according to claim 1, wherein, The causal reasoning algorithm uses the PC algorithm to construct a disaster causal graph based on historical event data, predicts the fault propagation path of topological mutation based on the causal graph, and generates a pre-intervention instruction set containing fault isolation, energy transfer and redundancy activation.

6. The method of optimal control of an electric power system according to claim 1, wherein The distributed learning framework is a federated learning architecture, each energy subsystem trains strategy model parameters locally, and uploads encrypted strategy model parameters to the parameter aggregation server regularly; The parameter aggregation server aggregates parameters from multiple subsystems through federated averaging algorithm to generate global model parameters; the global model parameters are processed by differential privacy mechanism, and the global model parameters are distributed to each energy subsystem for local model update.

7. The method of optimal control of an electric power system as claimed in claim 1, characterized in that, The multi-level response mechanism includes three levels of response: the first level of response performs fault isolation according to the causal reasoning within 0-5 minutes; the second level of response restores critical loads through edge node offline autonomy within 5-30 minutes; The third level of response is optimized through federated learning after communication recovery.

8. The method of optimal control of an electric power system of claim 1, wherein, The communication state judgment includes real-time monitoring of network delay and packet loss rate. When the delay exceeds the preset threshold and the packet loss rate exceeds the critical value, it is determined that the communication is interrupted, and the edge node automatically switches to the offline autonomous mode and calls the pre-stored scene strategy library to perform control actions. The scene strategy library contains pre-trained strategies based on historical event data through digital twin simulation, ensuring quick recovery of system functions when communication is interrupted.

9. The method of optimal control of an electric power system of claim 1, wherein, The scene strategy library contains pre-trained strategies for extreme scenarios, generated through digital twin environment simulation, including: constructing a digital twin environment, integrating historical extreme event data and multi-energy system physical models, simulating the dynamic behavior of the electric-thermal-gas coupled system, and introducing sensor data to generate extreme scenario sequences; In the digital twin environment, an improved SAC reinforcement learning algorithm is used for strategy training, and the effectiveness of the training results is verified. After verification, the compressed storage is stored in the local strategy library of the edge node, and is classified and labeled as a specific extreme scenario.

10. The method of optimal control of an electric power system according to claim 9, characterized in that, The verification of the effectiveness of the training results includes: verifying the trained strategy, and when the strategy meets the predefined performance indicators, it is compressed and stored in the local strategy library of the edge node; When the communication is interrupted, the edge node calls the matching pre-trained strategy from the local strategy library to perform control actions according to the current state; The performance indicators include a recovery time reduction rate of no less than 50%, and a key load guarantee rate of no less than 95%.

Images