Data-driven decision support system for power grid resilience enhancement

Abstract

The present application relates to the field of data acquisition control system, and discloses a power grid resilience improvement decision support system based on data driving.The system comprises a real-time data acquisition front end, which continuously acquires power grid node electrical quantity and distributed power output data; a multi-agent strategy interaction engine, which maps the distributed power and associated energy storage into independent agents, constructs a local observation space based on input data, and outputs inverter reference value adjustment and switch opening and closing state according to the communication and interaction of electrical topology and the execution of deep reinforcement learning strategy of the independent agents; and an island dynamic reconstruction control terminal, which dynamically adjusts the island boundary according to the opening and closing state and issues control signals to the inverter and circuit breaker according to the adjustment amount when the main network loses power.The present application realizes the cooperation of island boundary dynamic reconstruction and inverter bottom layer control, overcomes the mismatch defect of static island division under disaster and source and load fluctuation, maintains the stability of island voltage and frequency, and reduces the load shedding scale.

Description

Technical Field

[0001] This invention relates to the field of data acquisition and control systems, and discloses a data-driven decision support system for improving power grid resilience. Background Technology

[0002] When existing distribution networks experience power outages due to extreme disasters affecting the main grid, they typically rely on pre-defined static islanding schemes to ensure power supply to critical loads in specific areas. This conventional approach involves using deterministic optimization algorithms to determine a fixed island boundary and switching strategy offline, based on the initial topology of the distribution network, the rated capacity of distributed generation sources, and the priority labels of the loads. When a disaster occurs, the data acquisition and control system detects a main grid outage signal, directly retrieves and executes a pre-stored static switching sequence, isolating the area containing distributed generation sources from the main grid to form an island. During islanding operation, the distributed generation inverters employ pre-defined fixed droop control strategies or voltage source control strategies to maintain voltage and frequency within the island.

[0003] During islanded operation, conventional technical solutions are relatively passive in handling power imbalances within the island. When distributed power output fluctuates or loads change abruptly, existing systems use the proportional-integral regulator of the local controller to perform error-free tracking and adjustment based on the static deviations of voltage and frequency. If the power deficit exceeds the adjustment range, secondary loads are sequentially disconnected according to a predetermined load priority order until power balance is achieved. For line disconnections or equipment damage caused by the continuous evolution of disasters, existing systems typically wait until the fault is completely isolated and the network topology stabilizes before the centralized control center collects data from the entire network, recalculates offline to generate a new islanding scheme, and issues it for execution. During this process, the island boundaries remain fixed, and the inverter control mode and the switch topology state are independent of each other.

[0004] Existing static islanding schemes suffer from a spatiotemporal mismatch between the statically fixed island boundaries and the dynamic fluctuations of source and load under extreme disaster scenarios, leading to islanding collapse. Disaster evolution causes continuous changes in the distribution network topology, with source and load power exhibiting strong randomness and spatiotemporal coupling characteristics. Fixed island boundaries are highly susceptible to causing severe internal power imbalances. Islanding is decoupled from the inverter's underlying control. When topology changes abruptly or source and load fluctuate drastically, the response speed of the local proportional-integral regulator cannot match the transient impact. Power imbalances cause voltage over-limits and frequency deviations to propagate rapidly, leading to inverter disconnection and ultimately causing the entire islanding system to collapse. Summary of the Invention

[0005] The purpose of this invention is to provide a data-driven decision support system for improving power grid resilience, which can solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A data-driven decision support system for improving grid resilience includes: a real-time data acquisition front-end for continuously acquiring data on distribution network node voltage, frequency, line power, and distributed power generation output; a multi-agent strategy interaction engine for mapping distributed power generation and associated energy storage as independent agents, constructing a local observation space based on the data input from the real-time data acquisition front-end, wherein the independent agents communicate and interact according to electrical topology connections and execute deep reinforcement learning strategies, outputting the active and reactive power reference value adjustment amounts and tie switch opening and closing states of each distributed power generation inverter; and an islanding dynamic reconfiguration control terminal for summarizing the instructions output by the multi-agent strategy interaction engine, dynamically adjusting the islanding boundary according to the opening and closing states when the main grid loses power, and issuing control signals to the inverters and circuit breakers according to the active and reactive power reference value adjustment amounts.

[0007] Preferably, the local observation space includes the current output status of the distributed power source corresponding to the independent agent, the voltage amplitude and frequency offset of adjacent nodes, and the active and reactive power of adjacent lines; the multi-agent policy interaction engine also calculates the electrical distance and voltage phase angle difference between the independent agent and adjacent independent agents, and uses the electrical distance and voltage phase angle difference as supplementary feature dimensions of the local observation space to characterize the dynamic topological coupling strength of the independent agent in the distribution network. The local observation space is synchronously refreshed with the data update of the real-time data acquisition front end.

[0008] Preferably, the communication interaction between the independent intelligent agents is implemented based on a dynamic graph attention network. The dynamic graph attention network uses the independent intelligent agents as nodes and the electrical topology connection relationship as edges. It dynamically calculates the attention weight coefficient based on the electrical distance and voltage phase angle difference. The attention weight coefficient is updated as the operating state of the distribution network changes. The independent intelligent agents perform weighted aggregation of the observation features of neighboring independent intelligent agents through the dynamic graph attention network. The aggregated global features are then concatenated with the agent's own local observation space features and input into the strategy network to output the active and reactive power reference value adjustment amount and the tie switch opening and closing status.

[0009] Preferably, the multi-agent policy interaction engine is configured with a hybrid reward function when executing the deep reinforcement learning policy. The hybrid reward function consists of a voltage offset penalty term, a frequency deviation penalty term, and a critical load recovery reward term. The voltage offset penalty term is calculated based on the square of the difference between the node voltage amplitude obtained by the real-time data acquisition front end and the rated voltage. The frequency deviation penalty term is calculated based on the absolute value of the difference between the node frequency and the rated frequency. The critical load recovery reward term is calculated based on the power supply recovery status of the critical load after the islanded dynamic reconfiguration control terminal performs a switching action. The hybrid reward function guides the policy network to output actions that satisfy electrical constraints.

[0010] Preferably, when the islanded dynamic reconfiguration control terminal summarizes the instructions output by the multi-agent strategy interaction engine, it performs action mask verification on the opening and closing status of the tie switch based on the real-time topology connection relationship of the distribution network; the action mask verification identifies illegal switch actions that lead to ring network closure or disconnection of energized branches, and masks the illegal switch actions in the action space of the independent agents; the islanded dynamic reconfiguration control terminal only integrates the legal switch actions verified by the action mask with the active and reactive power reference value adjustment into a control instruction set, and sends it to the corresponding inverter and circuit breaker according to a preset time interval.

[0011] Preferably, when the real-time data acquisition front-end detects a branch disconnection fault in the distribution network, the multi-agent policy interaction engine removes the corresponding edge connection in the dynamic graph attention network and sets the feature vector of the disconnected node to zero. In the communication interaction of the independent agent in the next time step, the weighted aggregation is performed only based on the updated connected subgraph. When the communication link between the independent agent and the original adjacent independent agent is interrupted, the missing aggregation features are filled with historical communication features to maintain the stability of the input dimension and the continuous decision-making capability of the policy network in the topology change scenario.

[0012] Preferably, the voltage offset penalty term, the frequency deviation penalty term, and the critical load restoration reward term in the hybrid reward function are all configured with a time-series discount factor that decays exponentially with the time step. The multi-agent strategy interaction engine dynamically adjusts the time-series discount factor according to the operating conditions of the distribution network. When the distribution network is detected to be in the post-disaster steady-state recovery stage, the time-series discount factor is increased to strengthen the weight of long-term power supply restoration benefits. When the distribution network is detected to be in the fault transient impact stage, the time-series discount factor is decreased to prompt the strategy network to prioritize the output of control actions to suppress transient fluctuations.

[0013] Preferably, when the islanded dynamic reconfiguration control terminal executes the switching action corresponding to the open / closed state of the tie switch, it simultaneously sends a control mode switching command to the distributed power inverter in the target island. The control mode switching command controls the distributed power inverter to switch from grid following mode to islanded voltage source mode. During the switching transition, a virtual impedance voltage drop compensation component is generated based on the active and reactive power reference value adjustment. The virtual impedance voltage drop compensation component is superimposed on the voltage reference value of the islanded voltage source mode to suppress the transient current surge caused by the coupling between islanded boundary reconfiguration and control mode switching.

[0014] Preferably, it also includes a physical boundary verification terminal, which is pre-set with distribution network power flow constraint equations and equipment capacity limits; the physical boundary verification terminal intercepts the control command set sent by the islanded dynamic reconfiguration control terminal to the inverter and circuit breaker, and substitutes the control command set into the distribution network power flow constraint equations for over-limit verification; when the verification result indicates that there is a node voltage over-limit or branch power over-limit, the physical boundary verification terminal intercepts the over-limit command and sends a negative feedback penalty signal to the multi-agent strategy interaction engine, and at the same time triggers a preset local droop control strategy as a backup control command to be sent to the over-limit associated device.

[0015] Preferably, it also includes an experience playback management terminal, which stores transfer data generated after the independent agent performs an action. The transfer data includes the local observation space, the active and reactive power reference value adjustment amount and the opening and closing state of the interconnection switch, the hybrid reward function, and the local observation space of the next time step. The experience playback management terminal calculates the time-series difference error of the transfer data, prioritizes the transfer data according to the time-series difference error, and when the multi-agent policy interaction engine performs policy update sampling, it prioritizes extracting the high-priority transfer data and removes redundant data whose time-series difference error is lower than a preset threshold.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention maps distributed power sources and associated energy storage as independent intelligent agents. Based on electrical quantities input from the real-time data acquisition front-end, a local observation space is constructed. A deep reinforcement learning strategy is executed through a multi-agent policy interaction engine, outputting real-time inverter reference value adjustments and interconnection switch opening / closing states. This achieves dynamic reconfiguration of the islanding boundary as disasters evolve and load fluctuations occur. The multi-agent policy interaction engine communicates based on electrical topology connections. A dynamic graph attention network calculates attention weight coefficients based on electrical distance and voltage phase angle differences for feature weighting and aggregation, enabling each independent agent to perceive topological coupling strength and neighborhood states, and output coordinated control commands. The hybrid reward function consists of voltage offset penalty, frequency deviation penalty, and critical load recovery reward, guiding the policy network to output actions that satisfy electrical constraints. When adjusting the islanding boundary, the islanding dynamic reconfiguration control terminal simultaneously issues control mode switching commands and superimposes virtual impedance voltage drop compensation components, suppressing transient current surges caused by topology reconfiguration and mode switching coupling. This ensures the continuous and stable operation of the island under disturbances and overcomes the collapse defects caused by static boundaries and transient instability.

[0017] 2. This invention identifies and masks illegal switching actions that could lead to ring network closure or energized branch disconnection through action mask verification, ensuring the executability and security of issued control commands at the physical topology level. When a branch disconnection fault occurs in the distribution network, the dynamic graph attention network removes the corresponding edge connection and fills the missing aggregation features using historical communication characteristics, maintaining the stability of the input dimension and continuous decision-making capability of the strategy network under topology change scenarios. The time-series discount factor is dynamically adjusted according to operating conditions, increasing during the post-disaster steady-state recovery phase to enhance long-term power supply recovery benefits and decreasing during the fault transient impact phase to prioritize the output of actions that suppress transient fluctuations. The physical boundary verification terminal substitutes the control command set into the distribution network power flow constraint equations for over-limit verification, intercepts over-limit commands, and triggers the local droop control strategy, forming a physical security defense line. The experience playback management terminal prioritizes and samples the transfer data based on the time-series differential error and removes redundant data, improving the update efficiency and convergence stability of the strategy network. Attached Figure Description

[0018] Figure 1 Overall operation flowchart of the decision support system for improving power grid resilience; Figure 2 A schematic diagram illustrating the process of establishing and real-time updating the local observation space for independent intelligent agents; Figure 3 This is a schematic diagram illustrating the feature interaction aggregation process between agents in a dynamic graph attention network. Figure 4 A schematic diagram illustrating the process of constructing a multi-objective hybrid reward function and adjusting the discount factor for operating conditions. Figure 5A schematic diagram illustrating the process for verifying the operation of the interconnecting switch and controlling the switching of the inverter's operating mode in the islanded area; Figure 6 A schematic diagram illustrating the process for controlling instruction flow exceeding limits, verifying data usage, and optimizing management of reinforcement learning experience data. Detailed Implementation

[0019] refer to Figure 1 In one embodiment, the data-driven power grid resilience enhancement decision support system includes a real-time data acquisition front-end, a multi-agent strategy interaction engine, and an islanded dynamic reconfiguration control terminal. The real-time data acquisition front-end is deployed at key nodes of the distribution network. It continuously acquires data on voltage, frequency, line power, and distributed power output of distribution network nodes by establishing communication connections with local controllers of voltage transformers, current transformers, frequency transmitters, and distributed power sources in the distribution network. The real-time data acquisition front-end employs a time synchronization mechanism to ensure that the timestamp error of all acquired data does not exceed a preset threshold. The acquired data is filtered and then packaged and sent to the multi-agent strategy interaction engine according to a fixed time step. After receiving the data from the real-time data acquisition front-end, the multi-agent strategy interaction engine maps each distributed power source and its associated energy storage device in the distribution network as an independent agent, with each independent agent corresponding to a physical node in the distribution network. Based on the data input from the real-time data acquisition front-end, the multi-agent strategy interaction engine constructs a local observation space for each independent agent. This local observation space contains distribution network operating status information that the independent agent can directly observe. Independent agents communicate and interact with neighboring independent agents based on the electrical topology of the distribution network, exchanging their respective local observation information. The multi-agent policy interaction engine configures a deep reinforcement learning strategy for each independent agent. The independent agent fuses its own local observation information with communication information obtained from neighboring independent agents, inputs this fusion into the deep reinforcement learning strategy network, and outputs the active and reactive power reference value adjustments and tie switch opening / closing states for each distributed power inverter. The islanding dynamic reconfiguration control terminal establishes a communication connection with the multi-agent policy interaction engine, aggregating all instructions output by the engine. When the real-time data acquisition front-end detects a power outage in the backbone network, the islanding dynamic reconfiguration control terminal dynamically adjusts the island boundaries of the distribution network based on the tie switch opening / closing states output by the multi-agent policy interaction engine, dividing the distribution network into multiple independently operable island regions. Meanwhile, the island dynamic reconfiguration control terminal generates corresponding control signals based on the active and reactive power reference values ​​output by the multi-agent strategy interaction engine, and sends them to the circuit breakers corresponding to the distributed power inverters and interconnection switches in each island area to control the output power of the inverters and the opening and closing status of the circuit breakers.

[0020] In this embodiment, the real-time data acquisition front-end communicates with various measurement and control devices in the power distribution network via industrial communication protocols such as Modbus-RTU and IEC61850. For node voltage and frequency data, the front-end acquires data every 10 milliseconds; for line power and distributed power output data, it acquires data every 20 milliseconds. The acquired raw data is first processed by a sliding window filter to remove high-frequency noise interference. The sliding window size is set to 5 sampling points, and the average value within the window is taken as the valid data for the current moment. The filtered data is packaged according to a time step, with each data packet containing the voltage, frequency, line power, and distributed power output data of all nodes at the current moment. The time step of the data packet is set to 100 milliseconds. The real-time data acquisition front-end sends the packaged data to the multi-agent policy interaction engine via Ethernet. The data transmission process is encrypted to ensure data security and integrity.

[0021] The multi-agent policy interaction engine runs on an edge computing server deployed in the regional control center of the power distribution network. During the initialization phase, the engine creates an equal number of independent agent instances based on the power distribution network's topology. Each independent agent instance includes a local observation space construction module, a communication interaction module, and a policy execution module. The local observation space construction module receives data from the real-time data acquisition front-end, extracts operational status information relevant to the independent agent, and constructs the local observation space. The communication interaction module establishes communication connections with the communication interaction modules of adjacent independent agents based on the electrical topology of the power distribution network, exchanging local observation information. The policy execution module loads a pre-trained deep reinforcement learning policy network, inputs the fused observation information into the policy network, and outputs control actions.

[0022] The deep reinforcement learning policy network adopts an actor-critic architecture, consisting of a policy network and a value network. The policy network takes the fused observation information as input and outputs the active and reactive power reference value adjustments and tie switch opening / closing states of each distributed power inverter. The value network takes the same input as the policy network and outputs the value estimate of the current state. Both the policy and value networks use a fully connected neural network structure. The number of neurons in the input layer is equal to the dimension of the fused observation information. The hidden layer contains two fully connected layers, each with 128 neurons, and the ReLU activation function is used. The output layer of the policy network is divided into two parts: one part outputs the active and reactive power reference value adjustments using the tanh activation function, with an output range of [-1, 1], which is mapped to the actual adjustment range after linear transformation; the other part outputs the tie switch opening / closing states using the softmax activation function, outputting the probability distribution of each tie switch opening / closing state. The output layer of the value network contains one neuron, outputting the value estimate of the current state.

[0023] During the training phase, the multi-agent policy interaction engine employs a centralized training and distributed execution framework. All independent agents share a global value network, while each agent has its own independent policy network. During training, each agent collects its own transfer data, including its local observation space, executed actions, obtained rewards, and the local observation space for the next time step. The transfer data collected by all agents is aggregated into an experience replay buffer, which stores a certain amount of transfer data. At each policy update, a batch of transfer data is randomly selected from the experience replay buffer to update the global value network and the policy networks of each agent. The value network update uses a mean squared error loss function to minimize the error between the value estimate and the actual reward. The policy network update uses a policy gradient method to maximize the expected reward.

[0024] During the execution phase, each independent agent independently executes the policy network and outputs control actions based solely on its own local observations and communication information obtained from neighboring independent agents. Independent agents do not need to exchange global information; they only need to communicate locally with neighboring agents, reducing system communication overhead and computational complexity. Furthermore, the distributed execution approach makes the system more robust; if one independent agent fails, it will not affect the normal operation of other independent agents.

[0025] The islanded dynamic reconfiguration control terminal also runs on an edge computing server, sharing hardware resources with the multi-agent policy interaction engine. The islanded dynamic reconfiguration control terminal receives all control commands output by the multi-agent policy interaction engine, including the active and reactive power reference value adjustments of each distributed power inverter and the open / closed status of tie switches. The islanded dynamic reconfiguration control terminal summarizes and organizes the received control commands, distributing them to the corresponding islanded areas according to the distribution network topology information. When the real-time data acquisition front-end detects that the voltage of the main network is lower than a preset threshold and the duration exceeds a preset time, it determines that the main network is de-energized. At this time, the islanded dynamic reconfiguration control terminal generates circuit breaker control signals based on the tie switch open / closed status output by the multi-agent policy interaction engine, and sends them to the corresponding tie switch circuit breakers to control their opening and closing, dividing the distribution network into multiple islanded areas. Each islanded area contains at least one distributed power source and a certain number of loads, ensuring that the power within the islanded area can be basically balanced.

[0026] During islanded operation, the island dynamic reconfiguration control terminal continuously receives active and reactive power reference value adjustments from the multi-agent strategy interaction engine, generates inverter control signals, and distributes them to the distributed power inverters within each islanded area. The inverters adjust their output active and reactive power according to the received control signals to maintain voltage and frequency stability within the island. When the operating status of the distribution network changes, such as fluctuations in distributed power output, sudden load changes, or line faults, the multi-agent strategy interaction engine updates control commands in real time. The island dynamic reconfiguration control terminal dynamically adjusts the island boundaries and inverter output power based on the updated control commands to ensure the continuous and stable operation of the island.

[0027] refer to Figure 2In one embodiment, the local observation space includes the current output status of the distributed power source corresponding to the independent agent, the voltage amplitude and frequency offset of adjacent nodes, and the active and reactive power of adjacent lines. The current output status of the distributed power source includes the active and reactive power output of the distributed power source, in megawatts and megavars, respectively. Adjacent nodes refer to nodes directly connected to the node corresponding to the independent agent via lines, and the voltage amplitude of an adjacent node is the effective value of the line voltage of that node, in kilovolts. The frequency offset is the difference between the actual frequency and the rated frequency of the adjacent node, in hertz. The active and reactive power of adjacent lines are the active and reactive power transmitted on that line, in megawatts and megavars, respectively. The multi-agent policy interaction engine also calculates the electrical distance and voltage phase angle difference between the independent agent and adjacent independent agents, using the electrical distance and voltage phase angle difference as supplementary feature dimensions of the local observation space to characterize the dynamic topological coupling strength of the independent agents in the distribution network. The local observation space is refreshed synchronously with the data updates from the real-time data acquisition front end, and the refresh frequency is the same as the data transmission frequency of the real-time data acquisition front end.

[0028] Specifically, electrical distance measures the degree of electrical coupling between two nodes; the smaller the electrical distance, the stronger the electrical coupling between the two nodes. The formula for calculating electrical distance is as follows:

[0029] in, Represents a node With nodes Electrical distance between them Represents the th node in the admittance matrix Line number The modulus of each element in the column. The node admittance matrix is ​​a matrix describing the topology and electrical parameters of a distribution network; the elements in the matrix... Represents a node With nodes Mutual admittance between nodes, when nodes With nodes When there is no direct connection between them, .

[0030] The voltage phase angle difference is the difference between the voltage phase angles of two nodes, measured in radians. It reflects the power transfer between the two nodes; when active power is transferred between them, the voltage phase angle difference is not zero. The formula for calculating the voltage phase angle difference is as follows:

[0031] in, Represents a node With nodes The voltage phase angle difference between them Represents a node voltage phase angle, Represents a node The voltage phase angle.

[0032] When constructing the local observation space, the multi-agent policy interaction engine first extracts the active and reactive power outputs of the distributed power source corresponding to the independent agent from the data input from the real-time data acquisition front-end. Then, it extracts the voltage amplitude and frequency offset of all adjacent nodes, as well as the active and reactive power of all adjacent lines. Next, based on the current node admittance matrix and node voltage phase angle, the multi-agent policy interaction engine calculates the electrical distance and voltage phase angle difference between the independent agent and each adjacent independent agent. Finally, all the above features are concatenated into a vector in a preset order, serving as the local observation space for the independent agent. The feature dimensions of the local observation space are shown in Table 1.

[0033] Table 1. Dimensions of Local Observation Spatial Features In this embodiment, the dimension of the local observation space is . For example, for an independent agent with 3 neighboring nodes and 3 neighboring lines, the dimension of its local observation space is... Each feature in the local observation space is normalized, mapping the feature value to the [0,1] interval to improve the training efficiency and convergence speed of the deep reinforcement learning policy network. The normalization process uses the min-max normalization method, and the calculation formula is as follows:

[0034] in, These are the original eigenvalues. This is the minimum value of the feature. The maximum value of this feature. These are the normalized eigenvalues.

[0035] Each time the multi-agent policy interaction engine receives new data from the real-time data acquisition front-end, it recalculates the local observation space of each independent agent to ensure that the local observation space can reflect the operating status of the distribution network in real time. When the topology of the distribution network changes, such as when a line is disconnected or a switch is activated, the multi-agent policy interaction engine updates the node admittance matrix in real time and recalculates the electrical distance, updating the coupling feature dimension in the local observation space.

[0036] refer to Figure 3In one embodiment, communication between independent agents is implemented based on a dynamic graph attention network. The dynamic graph attention network constructs a dynamic graph structure with independent agents as nodes and the electrical topology connections of the distribution network as edges. The network dynamically calculates attention weight coefficients based on the electrical distance and voltage phase angle difference between each independent agent and its neighbors, and these coefficients are updated according to the operating state of the distribution network. Independent agents use the dynamic graph attention network to weighted aggregate the observation features of their neighbors, and then concatenate the aggregated global features with their own local observation spatial features before inputting them into the policy network to output active and reactive power reference value adjustments and the open / closed status of tie switches.

[0037] Specifically, the dynamic graph attention network comprises multiple graph attention layers, each transforming and aggregating the input node features. In this embodiment, the dynamic graph attention network includes two graph attention layers. The input of the first graph attention layer is the local observation space features of each independent agent, and the output is the node features after one feature transformation and aggregation. The input of the second graph attention layer is the output of the first graph attention layer, and the output is the global features after a second feature transformation and aggregation.

[0038] The computation process for each graph attention layer is as follows: First, a linear transformation is performed on the input features of each node to obtain the transformed feature vector. The parameters of the linear transformation are a learnable weight matrix, which is updated during training using the backpropagation algorithm. Then, the attention weight coefficients between each node and its neighboring nodes are calculated. The formula for calculating the attention weight coefficients is as follows:

[0039] in, Represents a node For nodes Attention weight coefficient, Represents a node The input feature vector, Represents a node The input feature vector, The weight matrix represents the linear transformation. The weight vector representing the attention mechanism. This represents a vector concatenation operation. Represents a node With nodes Electrical distance between them Represents a node With nodes The voltage phase angle difference between them Represents a node The set of adjacent nodes, This represents the linear rectified activation function with leakage.

[0040] The attention weight coefficients are related not only to the feature vectors of nodes, but also to the electrical distance and voltage phase angle difference between nodes. The smaller the electrical distance and the larger the voltage phase angle difference, the stronger the electrical coupling between nodes, and the larger the attention weight coefficients. In this way, the dynamic graph attention network can dynamically adjust the attention weights between nodes according to the operating state of the distribution network, making independent agents pay more attention to the state information of neighboring independent agents with strong electrical coupling.

[0041] After calculating the attention weight coefficients, the transformed feature vectors of adjacent nodes are weighted and aggregated to obtain the node... The output feature vector. The formula for weighted aggregation is as follows:

[0042] in, Represents a node The output feature vector, This represents the activation function; in this embodiment, the ReLU activation function is used.

[0043] After processing by two graph attention layers, each independent agent obtains a global feature vector that aggregates neighborhood information. This global feature vector is then concatenated with the independent agent's own local observation space feature vector to obtain a fused feature vector. This fused feature vector is input into the input layer of a deep reinforcement learning policy network. After processing by the policy network, the active and reactive power reference value adjustments and interconnection switch opening / closing states of each distributed power inverter are output. The action space of the independent agents is defined in Table 2.

[0044] Table 2 Definition of Action Space for Independent Agents In this embodiment, the action space of the independent intelligent agent comprises two parts: continuous actions and discrete actions. The continuous actions part includes active power reference value adjustment and reactive power reference value adjustment, used to adjust the output power of the distributed power inverter. The discrete actions part includes the open and closed states of tie switches, used to control the opening and closing of tie switch circuit breakers. For a distribution network with K tie switches, the dimension of the discrete action space is K, and each dimension corresponds to the open and closed state of one tie switch.

[0045] refer to Figure 4In one embodiment, the multi-agent policy interaction engine is configured with a hybrid reward function when executing deep reinforcement learning policies. The hybrid reward function consists of a voltage offset penalty term, a frequency deviation penalty term, and a critical load restoration reward term. The voltage offset penalty term is calculated based on the square of the difference between the node voltage amplitude obtained from the real-time data acquisition front-end and the rated voltage; the frequency deviation penalty term is calculated based on the absolute value of the difference between the node frequency and the rated frequency; and the critical load restoration reward term is calculated based on the power supply restoration status of the critical load after the islanded dynamic reconfiguration control terminal performs a switching action. The hybrid reward function guides the policy network to output actions that satisfy electrical constraints.

[0046] Specifically, the formula for calculating the hybrid reward function is as follows: in, This represents the value of the mixed reward function. This indicates a voltage offset penalty term. This indicates the frequency deviation penalty term. This indicates the critical load recovery reward item. , , These represent the weighting coefficients for the voltage offset penalty, frequency deviation penalty, and critical load recovery reward, respectively.

[0047] The voltage offset penalty term is used to penalize situations where the node voltage deviates from the rated voltage. The calculation formula is as follows: in, This indicates the number of nodes that need to be monitored. Represents a node The actual voltage amplitude, This indicates the rated voltage amplitude of the power distribution network.

[0048] The frequency deviation penalty term is used to penalize situations where the node frequency deviates from the rated frequency. The calculation formula is as follows: in, This indicates the number of nodes that need to be monitored. Represents a node The actual frequency, This indicates the rated frequency of the power distribution network.

[0049] The critical load restoration incentive is used to reward the restoration of power supply to critical loads. The calculation formula is as follows: in, Indicates the number of critical loads. Indicates critical load Power supply status, Indicates critical load Powered Indicates critical load The power is out. Indicates critical load Rated power.

[0050] In this embodiment, the weighting coefficient , , The values ​​are 1.0, 1.0, and 1.0 respectively. During training, the weight coefficients can be adjusted according to actual needs to balance the importance of different objectives. For example, if more attention is paid to voltage stability, the weights can be increased. The value of ; if more attention is paid to the power restoration of critical loads, it can be increased The value of .

[0051] The hybrid reward function combines the three objectives of voltage stability, frequency stability, and critical load restoration, guiding the policy network to output control actions that simultaneously satisfy these three objectives. When an independent agent performs an action, if the node voltage and frequency get closer to their rated values, or more critical loads are restored to power, the hybrid reward function value will increase, and the policy network will tend to perform similar actions in future states. Conversely, if the node voltage and frequency deviate more from their rated values, or more critical loads experience power outages, the hybrid reward function value will decrease, and the policy network will avoid performing similar actions in future states.

[0052] refer to Figure 5 In one embodiment, when the islanded dynamic reconfiguration control terminal aggregates the instructions output by the multi-agent strategy interaction engine, it performs action mask verification on the opening and closing status of tie switches based on the real-time topology connection relationship of the distribution network. The action mask verification identifies illegal switch actions that lead to ring network closure or disconnection of energized branches, and masks these illegal switch actions in the action space of the independent agents. The islanded dynamic reconfiguration control terminal only integrates the legal switch actions verified by the action mask with the active and reactive power reference value adjustments into a control instruction set, and issues it to the corresponding inverters and circuit breakers according to a preset timing interval.

[0053] Specifically, the islanded dynamic reconfiguration control terminal maintains a real-time topology model of the distribution network. This model includes the connection relationships of all nodes, lines, and switches in the distribution network, as well as their current operating status. When the multi-agent policy interaction engine outputs a tie switch opening / closing status command, the islanded dynamic reconfiguration control terminal first applies this command to the real-time topology model to simulate the distribution network topology after the switch action. Then, the islanded dynamic reconfiguration control terminal checks whether the simulated topology contains closed loops or disconnected energized branches.

[0054] A closed loop in a distribution network refers to the formation of a closed loop, which leads to circulating current and affects the safe and stable operation of the distribution network. The islanded dynamic reconfiguration control terminal uses a depth-first search algorithm to traverse the distribution network topology to check for loops. If a loop exists, the switch action is determined to be a violation.

[0055] Disconnecting a energized branch refers to cutting off a branch that is currently operating with power, which will cause a power outage to the loads connected to that branch, affecting power supply reliability. The islanded dynamic reconfiguration control terminal checks whether the branch corresponding to each switch action is in an energized state. If the branch is in an energized state, and disconnecting the branch will cause the loads connected to that branch to lose power, then the switch action is determined to be an illegal action.

[0056] For identified unauthorized switching actions, the islanded dynamic reconfiguration control terminal generates an action mask. The action mask is a vector with the same dimension as the discrete action space, where each element corresponds to the open / closed state of a switch. For unauthorized switching actions, the corresponding element in the action mask is set to 0, indicating that the action is blocked; for legitimate switching actions, the corresponding element is set to 1, indicating that the action is allowed. The islanded dynamic reconfiguration control terminal sends the action mask to the multi-agent policy interaction engine. When executing the policy network, the multi-agent policy interaction engine multiplies the action mask by the probability distribution of the discrete actions, sets the probability of blocked actions to 0, and then re-normalizes the probability distribution to ensure that the sum of the probabilities of all legitimate actions is 1.

[0057] After action mask verification, the islanded dynamic reconfiguration control terminal integrates legitimate switching actions with active and reactive power reference value adjustments into a control command set. The control command set includes the active and reactive power reference value adjustments for each distributed power inverter and the open / closed status of each tie switch. The islanded dynamic reconfiguration control terminal issues the control command set according to a preset timing interval, which is the same as the data transmission frequency of the real-time data acquisition front-end, at 100 milliseconds. For inverter control commands, the islanded dynamic reconfiguration control terminal directly issues them to the corresponding distributed power inverter; for circuit breaker control commands, the islanded dynamic reconfiguration control terminal first issues them to the corresponding circuit breaker controller, which then executes the switching action.

[0058] In one embodiment, when the real-time data acquisition front-end detects a branch disconnection fault in the distribution network, the multi-agent policy interaction engine removes the corresponding edge connection in the dynamic graph attention network and sets the feature vector of the disconnected node to zero. In the next time step's communication interaction, independent agents perform weighted aggregation only based on the updated connected subgraph. When the communication link between an independent agent and its former neighboring independent agents is interrupted, historical communication features are used to fill in the missing aggregation features to maintain the stability of the input dimension and the continuous decision-making capability of the policy network under topology change scenarios.

[0059] Specifically, the real-time data acquisition front-end detects branch disconnection faults by monitoring the current and voltage of the line. When the current in the line suddenly drops to zero and a significant difference appears in the voltage across the two ends of the line, it is determined that a branch disconnection fault has occurred. The real-time data acquisition front-end sends the branch disconnection fault information to the multi-agent policy interaction engine and the islanded dynamic reconfiguration control terminal.

[0060] Upon receiving a branch disconnection fault information, the multi-agent policy interaction engine first updates the distribution network topology information, removing the edge connections corresponding to the disconnected branch. Then, it updates the graph structure of the dynamic graph attention network, removing the corresponding edge connections. For the two end nodes of the disconnected branch, their feature vectors in the dynamic graph attention network are set to zero, indicating that there is no longer a communication connection between these two nodes.

[0061] In the next communication step, independent agents only communicate with neighboring independent agents in the updated connected subgraph. The dynamic graph attention network only performs weighted aggregation of node features in the connected subgraph. For independent agents whose communication links are interrupted due to branch breaks, they can no longer obtain the latest communication features from their original neighboring independent agents. At this time, the independent agent fills the missing aggregated features using historical communication features. Historical communication features refer to the last communication features obtained by the independent agent from its original neighboring independent agents before the communication link was interrupted. The independent agent performs weighted aggregation of historical communication features with the communication features currently obtained from other neighboring independent agents to obtain the aggregated global feature vector.

[0062] Filling in missing aggregation features with historical communication characteristics can maintain the stability of the policy network's input dimensions, preventing it from malfunctioning due to changes in the input dimensions caused by topological abrupt changes. Simultaneously, historical communication characteristics can reflect the operational status of previously adjacent independent agents to a certain extent, enabling these agents to make reasonable decisions even shortly after a communication link is interrupted, thus maintaining the system's continuous decision-making capability.

[0063] In one embodiment, the voltage offset penalty term, frequency deviation penalty term, and critical load restoration reward term in the hybrid reward function are all configured with a time-series discount factor that decays exponentially with the time step. The multi-agent policy interaction engine dynamically adjusts the time-series discount factor according to the operating conditions of the distribution network. When the distribution network is detected to be in the post-disaster steady-state recovery phase, the time-series discount factor is increased to strengthen the weight of long-term power supply restoration benefits; when the distribution network is detected to be in the fault transient impact phase, the time-series discount factor is decreased to prompt the policy network to prioritize outputting control actions that suppress transient fluctuations.

[0064] Specifically, the time-series discount factor is used to calculate the present value of future rewards, and its value ranges from [0,1]. A larger time-series discount factor results in a larger present value of future rewards, and the strategy network will focus more on long-term returns; a smaller time-series discount factor results in a smaller present value of future rewards, and the strategy network will focus more on short-term returns. The formula for calculating the time-series discount factor is as follows:

[0065] in, Indicates time The time-series discount factor, Represents the initial time-series discount factor. Indicates the attenuation coefficient. Indicates the time step.

[0066] The multi-agent strategy interaction engine determines the operating condition of the distribution network based on the real-time data acquisition front-end information. The operating condition is divided into two stages: the fault transient impact stage and the post-disaster steady-state recovery stage. The fault transient impact stage refers to the period immediately following a fault in the distribution network, during which voltage and frequency fluctuate drastically; the post-disaster steady-state recovery stage refers to the period when the voltage and frequency of the distribution network have basically stabilized, and power supply to critical loads is gradually being restored.

[0067] When the distribution network is detected to be in a fault transient impact phase, the multi-agent policy interaction engine reduces the initial time-series discount factor. The value of is adjusted, and the attenuation coefficient is increased. The value of causes the time-series discount factor to decay rapidly. This means the policy network will focus more on short-term gains, prioritizing control actions that can quickly suppress voltage and frequency fluctuations, thus preventing distribution network collapse.

[0068] When the distribution network is detected to be in the post-disaster steady-state recovery phase, the multi-agent policy interaction engine increases the initial time-series discount factor. The value of is determined while reducing the attenuation coefficient. The value of causes the time-series discount factor to decay slowly. This allows the policy network to focus more on long-term gains, prioritizing control actions that restore power to more critical loads and improving the power supply recovery capability of the distribution network. Refer to Table 3 for the time-series discount factor configuration under different operating conditions.

[0069] Table 3. Configuration of Time Series Discount Factors under Different Operating Conditions In this embodiment, the multi-agent policy interaction engine determines the operating condition of the distribution network by monitoring the fluctuation amplitude of node voltage and frequency. When the fluctuation amplitude of node voltage exceeds 10% of the rated voltage, or the fluctuation amplitude of node frequency exceeds 0.5Hz of the rated frequency, the distribution network is determined to be in the fault transient impact stage. When the fluctuation amplitude of node voltage is less than 5% of the rated voltage, and the fluctuation amplitude of node frequency is less than 0.2Hz of the rated frequency, and the duration exceeds 10 seconds, the distribution network is determined to be in the post-disaster steady-state recovery stage.

[0070] In one embodiment, when the islanded dynamic reconfiguration control terminal executes the switching action corresponding to the open / closed state of the tie switch, it simultaneously issues a control mode switching command to the distributed power inverter within the target island. The control mode switching command controls the distributed power inverter to switch from grid-following mode to islanded voltage source mode. During the switching transition, a virtual impedance voltage drop compensation component is generated based on the active and reactive power reference value adjustment. This virtual impedance voltage drop compensation component is superimposed on the voltage reference value of the islanded voltage source mode to suppress the transient current surge caused by the coupling between islanded boundary reconfiguration and control mode switching.

[0071] Specifically, distributed generation inverters operate in grid-following mode, where the inverter's output current tracks the phase and frequency of the grid voltage, and the output power is determined by active and reactive power reference values. When the main grid loses power and the distribution network enters islanded operation mode, the distributed generation inverter needs to switch to islanded voltage source mode, where the inverter acts as a voltage source to provide stable voltage and frequency to the island.

[0072] The control mode switching process is dynamic. Improper control during this process can lead to significant transient surges in the inverter's output current, potentially damaging the inverter equipment. Simultaneously, the islanding boundary reconfiguration process alters the distribution network topology, causing voltage and frequency fluctuations. The coupling effect between islanding boundary reconfiguration and control mode switching further exacerbates transient shocks, impacting the stable operation of the islanded network.

[0073] To suppress transient current surges, the islanded dynamic reconfiguration control terminal, while issuing control mode switching commands, generates a virtual impedance voltage drop compensation component based on the active and reactive power reference values ​​output by the multi-agent strategy interaction engine. The calculation formula for the virtual impedance voltage drop compensation component is as follows:

[0074] in, This represents the virtual impedance voltage drop compensation component. Indicates virtual resistance. Represents virtual reactance. This represents the reference value of the inverter output current, which is calculated from the active and reactive power reference value adjustment.

[0075] The virtual impedance voltage drop compensation component simulates the voltage drop across the inverter's output impedance. When superimposed on the voltage reference value in islanded voltage source mode, it can compensate for the voltage drop caused by changes in the inverter's output current, thus keeping the inverter's output voltage stable during control mode switching and suppressing transient current surges.

[0076] During the control mode switching transition, the islanded dynamic reconfiguration control terminal gradually increases the amplitude of the virtual impedance drop compensation component, while gradually decreasing the current reference value in grid follower mode and gradually increasing the voltage reference value in islanded voltage source mode. The transition time is set to 100 milliseconds to ensure a smooth control mode switching process. After the transition process is completed, the distributed power inverter fully enters the islanded voltage source mode, and the virtual impedance drop compensation component remains unchanged, continuing to provide stable voltage and frequency to the island.

[0077] refer to Figure 6 In one embodiment, the data-driven power grid resilience enhancement decision support system further includes a physical boundary verification terminal. The physical boundary verification terminal is pre-configured with distribution network power flow constraint equations and equipment capacity limits. The physical boundary verification terminal intercepts the control command set sent from the islanded dynamic reconfiguration control terminal to the inverters and circuit breakers, and substitutes the control command set into the distribution network power flow constraint equations for limit exceedance verification. When the verification result indicates that there is a node voltage exceedance or branch power exceedance, the physical boundary verification terminal intercepts the limit exceedance command and sends a negative feedback penalty signal to the multi-agent strategy interaction engine. Simultaneously, it triggers a preset local droop control strategy as a backup control command to send to the related limit exceedance devices.

[0078] Specifically, the power flow constraint equations of the distribution network describe the relationship between node voltages, injected power, and branch power in the distribution network. In this embodiment, the Newton-Raphson method is used to solve the power flow equations of the distribution network. The physical boundary verification terminal is pre-set with the node admittance matrix, node load data, and equipment capacity limits of the distribution network. The equipment capacity limits include the active and reactive power capacity limits of distributed generation inverters, the active and reactive power transmission capacity limits of lines, and the capacity limits of transformers.

[0079] After the islanded dynamic reconfiguration control terminal generates a control command set, the physical boundary verification terminal first intercepts this control command set. Then, based on the open / closed status of the tie switches in the control command set, the physical boundary verification terminal updates the distribution network topology and node admittance matrix. Next, the physical boundary verification terminal uses the active and reactive power reference value adjustments in the control command set as the injected power of distributed generation, substitutes them into the distribution network power flow constraint equations, and solves for the power flow distribution of the distribution network.

[0080] After obtaining the power flow distribution, the physical boundary verification terminal checks whether the voltage amplitude of each node is within the allowable range and whether the transmission power of each branch exceeds the equipment capacity limit. The allowable range for node voltage is 90% to 110% of the rated voltage. If the node voltage is lower than 90% or higher than 110% of the rated voltage, it is determined that the node voltage is out of limit. If the transmission power of a branch exceeds the equipment capacity limit, it is determined that the branch power is out of limit.

[0081] If the verification result indicates that a node voltage or branch power exceeds the limit, the physical boundary verification terminal intercepts the corresponding limit-exceeding instruction and does not send it to the device. Simultaneously, the physical boundary verification terminal sends a negative feedback penalty signal to the multi-agent policy interaction engine, with a value of -1.0. Upon receiving the negative feedback penalty signal, the multi-agent policy interaction engine adds it as an additional penalty to the hybrid reward function, guiding the policy network to avoid outputting similar limit-exceeding instructions in future decisions.

[0082] In addition, the physical boundary verification terminal triggers a preset local droop control strategy as a backup control command, which is then sent to the associated out-of-limit devices. The local droop control strategy is a distributed control strategy that does not rely on global information; it adjusts the device's output power solely based on local voltage and frequency information. For distributed power inverters, the calculation formula for the local droop control strategy is as follows:

[0083] in, and These represent the active power and reactive power output by the inverter, respectively. and These represent the reference values ​​for the active power and reactive power of the inverter, respectively. and These represent the active power-frequency droop coefficient and the reactive power-voltage droop coefficient, respectively. and These represent the local frequency and voltage, respectively. and These represent the rated frequency and rated voltage, respectively.

[0084] The local droop control strategy can quickly respond to changes in voltage and frequency, adjust the output power of equipment, suppress further deterioration of over-limit situations, and ensure the safe and stable operation of the distribution network. Physical boundary verification rules are detailed in Table 4.

[0085] Table 4 Physical Boundary Verification Rules In one embodiment, the data-driven power grid resilience enhancement decision support system further includes an experience playback management terminal. The experience playback management terminal stores transfer data generated after independent agents execute actions. This transfer data includes local observation space, active and reactive power reference value adjustments and tie switch opening / closing states, a hybrid reward function, and the local observation space for the next time step. The experience playback management terminal calculates the time-series difference error of the transfer data and prioritizes the transfer data based on this error. When the multi-agent policy interaction engine performs policy update sampling, it prioritizes extracting high-priority transfer data and removes redundant data with time-series difference errors below a preset threshold.

[0086] Specifically, each independent agent generates a transition data entry after performing an action. The transition data entry is a quintuple. ,in Represents the local observation space at the current moment. This indicates the actions performed, including adjustments to active and reactive power reference values ​​and the open / closed status of the interconnecting switch. This represents the obtained mixed reward function value. Indicates the local observation space of the next time step. Indicates the end of the process. This indicates the end of the current round. This indicates that the current round continues.

[0087] The independent intelligent agent sends the generated transfer data to the experience playback management terminal. The experience playback management terminal maintains a priority experience playback buffer, which stores a certain number of transfer data points. Each transfer data point has a corresponding priority, and the priority value is proportional to the timing difference error. The formula for calculating the timing difference error is as follows:

[0088] in, Indicates timing difference error. This represents the value of the mixed reward function. Represents the time-series discount factor. This represents the estimated value of the state at the next time step. This represents the estimated value of the current state.

[0089] Temporal difference error reflects the accuracy of the value network's estimation of the current state's value. The larger the temporal difference error, the less accurate the value network's estimation of the current state's value. This transition data is more crucial for updating the value network and policy network, and therefore should be given higher priority.

[0090] After receiving the transferred data, the experience playback management terminal first calculates the timing difference error of the transferred data. Then, based on the magnitude of the timing difference error, it assigns a priority to the transferred data. The priority calculation formula is as follows:

[0091] in, Indicates the priority of data transfer. It represents a very small positive number, used to avoid cases where the priority is zero.

[0092] The experience playback management terminal stores the transferred data in a priority experience playback buffer according to its priority from high to low. When the buffer is full, the lowest priority transferred data is removed to make room for new transferred data.

[0093] During policy updates in the multi-agent policy interaction engine, the experience replay management terminal samples transition data based on its priority. The sampling probability is directly proportional to the priority of the transition data; higher-priority transition data has a greater probability of being sampled. In this way, the multi-agent policy interaction engine can prioritize the use of transition data that is more valuable for policy updates, improving the efficiency and convergence speed of policy updates.

[0094] In addition, the experience replay management terminal periodically checks the transfer data in the priority experience replay buffer and removes redundant data whose temporal difference error is below a preset threshold. Redundant data refers to transfer data for which the value network has already accurately estimated the state value. This data has little effect on policy updates, and removing it can save storage space and improve sampling efficiency. The preset threshold is adjusted according to the actual situation. In this embodiment, the preset threshold is set to 0.01.

[0095] In one embodiment, the workflow of a data-driven power grid resilience enhancement decision support system is as follows: A real-time data acquisition front-end continuously collects operational status data of the distribution network. After filtering and packaging, this data is sent to a multi-agent policy interaction engine. The multi-agent policy interaction engine maps distributed power sources and associated energy storage as independent agents, constructing a local observation space for each agent. Independent agents communicate and interact with neighboring agents through a dynamic graph attention network, aggregating neighborhood information. The independent agents input the fused observation information into a deep reinforcement learning policy network, outputting control actions. The islanded dynamic reconfiguration control terminal aggregates the control actions output by the multi-agent policy interaction engine, performs action mask verification, identifies and blocks illegal switching actions. When the main grid loses power, the islanded dynamic reconfiguration control terminal dynamically adjusts the island boundary based on legitimate switching actions, synchronously sends control mode switching commands to the distributed power source inverters, and generates virtual impedance voltage drop compensation components to suppress transient impacts. The physical boundary verification terminal intercepts the control command set, performs power flow over-limit verification, intercepts over-limit commands, and triggers a backup control strategy. The experience playback management terminal stores and transfers data, calculates time-series differential errors, performs priority sorting and sampling, and supports policy updates for the multi-agent policy interaction engine.

[0096] In this embodiment, the system can perceive the operating status of the distribution network in real time and dynamically adjust the output power of the island boundary and distributed power sources, realizing the dynamic reconstruction of the island boundary as disasters evolve and source-load fluctuations occur. The multi-agent policy interaction engine uses a dynamic graph attention network for communication and interaction, which can dynamically adjust the attention weight according to the operating status of the distribution network, enabling independent agents to perceive the topological coupling strength and neighborhood state, and output coordinated control commands. A hybrid reward function guides the policy network to output actions that meet electrical constraints, ensuring the voltage and frequency stability of the island. Action mask verification and physical boundary verification constitute a double security defense, ensuring that the issued control commands are safe and feasible at both the physical topology level and the electrical constraint level. The experience playback management terminal adopts a priority experience playback mechanism, which improves the update efficiency and convergence stability of the policy network. The system can effectively overcome the defects of island collapse caused by static boundaries and transient instability, ensuring the continuous and stable operation of the island under disturbances.

Claims

1. A data-driven decision support system for improving power grid resilience, characterized in that, include: The real-time data acquisition front-end is used to continuously acquire data on voltage, frequency, line power, and distributed power output of distribution network nodes; A multi-agent strategy interaction engine is used to map distributed power sources and associated energy storage as independent agents. Based on the data input from the real-time data acquisition front end, a local observation space is constructed. The independent agents communicate and interact according to the electrical topology connection relationship and execute deep reinforcement learning strategies to output the active and reactive power reference value adjustment amount and the interconnection switch opening and closing status of each distributed power source inverter. The island dynamic reconfiguration control terminal is used to summarize the instructions output by the multi-agent strategy interaction engine. When the backbone network loses power, it dynamically adjusts the island boundary according to the opening and closing state, and sends control signals to the inverter and circuit breaker according to the active and reactive power reference value adjustment.

2. The data-driven power grid resilience enhancement decision support system according to claim 1, characterized in that, The local observation space includes the current output status of the distributed power source corresponding to the independent intelligent agent, the voltage amplitude and frequency offset of adjacent nodes, and the active power and reactive power of adjacent lines. The multi-agent strategy interaction engine also calculates the electrical distance and voltage phase angle difference between the independent agent and its neighboring independent agents, and uses the electrical distance and voltage phase angle difference as a supplementary feature dimension of the local observation space to characterize the dynamic topological coupling strength of the independent agent in the power distribution network. The local observation space is synchronously refreshed with the data update of the real-time data acquisition front end.

3. The data-driven power grid resilience enhancement decision support system according to claim 2, characterized in that, The communication and interaction between the independent intelligent agents is realized based on a dynamic graph attention network. The dynamic graph attention network uses the independent intelligent agents as nodes and the electrical topology connection relationship as edges. It dynamically calculates the attention weight coefficient based on the electrical distance and voltage phase angle difference. The attention weight coefficient is updated as the operating state of the distribution network changes. The independent agent performs weighted aggregation of the observation features of neighboring independent agents through the dynamic graph attention network, and then inputs the aggregated global features and its own local observation space features into the policy network to output the active and reactive power reference value adjustment amount and the opening and closing status of the interconnection switch.

4. The data-driven power grid resilience enhancement decision support system according to claim 1, characterized in that, When the multi-agent policy interaction engine executes the deep reinforcement learning policy, it is configured with a hybrid reward function, which consists of a voltage offset penalty term, a frequency deviation penalty term, and a critical load recovery reward term. The voltage offset penalty term is calculated based on the square of the difference between the node voltage amplitude obtained by the real-time data acquisition front end and the rated voltage. The frequency deviation penalty term is calculated based on the absolute value of the difference between the node frequency and the rated frequency. The critical load restoration reward term is calculated based on the power supply restoration status of the critical load after the islanded dynamic reconfiguration control terminal performs a switching action. The hybrid reward function guides the strategy network to output actions that satisfy electrical constraints.

5. The data-driven power grid resilience enhancement decision support system according to claim 1, characterized in that, When the island dynamic reconfiguration control terminal summarizes the instructions output by the multi-agent strategy interaction engine, it performs action mask verification on the opening and closing status of the tie switch based on the real-time topology connection relationship of the distribution network. The action mask verification identifies illegal switch actions that cause the ring network to close or the energized branch to be disconnected, and shields the illegal switch actions in the action space of the independent intelligent agent. The islanded dynamic reconfiguration control terminal integrates the legitimate switching actions verified by the action mask with the active and reactive power reference value adjustment into a control command set, and sends it to the corresponding inverter and circuit breaker according to a preset timing interval.

6. The data-driven power grid resilience enhancement decision support system according to claim 3, characterized in that, When the real-time data acquisition front end detects a branch disconnection fault in the distribution network, the multi-agent strategy interaction engine removes the corresponding edge connection in the dynamic graph attention network and sets the feature vector of the disconnected node to zero. In the next communication interaction, the independent agent performs the weighted aggregation only based on the updated connected subgraph. When the communication link between the independent agent and its original neighboring independent agents is interrupted, the missing aggregation features are filled with historical communication features to maintain the stability of the input dimension and the continuous decision-making ability of the policy network in the topology change scenario.

7. The data-driven power grid resilience enhancement decision support system according to claim 4, characterized in that, The voltage offset penalty term, the frequency deviation penalty term, and the critical load recovery reward term in the hybrid reward function are all configured with a time-series discount factor that decays exponentially with the time step. The multi-agent strategy interaction engine dynamically adjusts the time-series discount factor according to the operating conditions of the distribution network. When the distribution network is detected to be in the post-disaster steady-state recovery stage, the time-series discount factor is increased to strengthen the weight of long-term power supply recovery benefits. When the distribution network is detected to be in a fault transient impact phase, the timing discount factor is reduced to prompt the strategy network to prioritize outputting control actions to suppress transient fluctuations.

8. The data-driven power grid resilience enhancement decision support system according to claim 5, characterized in that, When the island dynamic reconfiguration control terminal executes the switching action corresponding to the opening and closing state of the interconnection switch, it simultaneously sends a control mode switching command to the distributed power inverter in the target island. The control mode switching command controls the distributed power inverter to switch from grid follower mode to islanded voltage source mode. During the switching transition, a virtual impedance voltage drop compensation component is generated based on the active and reactive power reference value adjustment. The virtual impedance voltage drop compensation component is superimposed on the voltage reference value of the islanded voltage source mode to suppress the transient current impact caused by the coupling between the islanded boundary reconfiguration and the control mode switching.

9. The data-driven power grid resilience enhancement decision support system according to claim 1, characterized in that, It also includes a physical boundary verification terminal, which is pre-set with power flow constraint equations for the distribution network and equipment capacity limits; The physical boundary verification terminal intercepts the control command set sent by the islanded dynamic reconfiguration control terminal to the inverter and circuit breaker, and substitutes the control command set into the power flow constraint equation of the distribution network for over-limit verification. When the verification result indicates that there is a node voltage exceeding the limit or a branch power exceeding the limit, the physical boundary verification terminal intercepts the limit exceeding instruction and sends a negative feedback penalty signal to the multi-agent policy interaction engine. At the same time, it triggers a preset local droop control strategy as a backup control instruction to be sent to the limit-related device.

10. The data-driven power grid resilience enhancement decision support system according to claim 1, characterized in that, It also includes an experience playback management terminal, which stores the transfer data generated after the independent intelligent agent performs an action. The transfer data includes the local observation space, the active and reactive power reference value adjustment amount and the opening and closing state of the interconnection switch, the hybrid reward function and the local observation space of the next time step. The experience playback management terminal calculates the temporal difference error of the transfer data, prioritizes the transfer data according to the temporal difference error, and when the multi-agent policy interaction engine performs policy update sampling, it first extracts the transfer data with high priority and removes redundant data whose temporal difference error is lower than a preset threshold.

Images