Power distribution network voltage out-of-limit prevention control method and system based on reinforcement learning

Abstract

The application provides a power distribution network voltage out-of-limit prevention control method and system based on reinforcement learning, and relates to the technical field of electric power. The method comprises the following steps: based on interactive trajectory and reward and punishment feedback, multi-objective decomposition pre-training and random weight sampling are carried out by using meta-reinforcement learning to obtain a multi-objective Pareto frontier strategy solution set; the multi-objective Pareto frontier strategy solution set is deployed to a power distribution network real-time control system, voltage regulation device control instructions are matched and issued according to a current multi-dimensional state vector, and closed-loop iterative control is realized through state transition update and voltage monitoring. The application is based on data repair, dynamic boundary construction, double-agent game and meta-reinforcement learning, obtains a strategy solution set and realizes real-time deployment, and realizes closed-loop dynamic control of power distribution network voltage out-of-limit.

Description

Technical Field

[0001] This invention relates to the field of power technology, and in particular to a method and system for preventing and controlling voltage overruns in distribution networks based on reinforcement learning. Background Technology

[0002] With the high proportion of distributed renewable energy access, the power distribution network operation data exhibits strong spatiotemporal coupling and high-dimensional heterogeneity. The accuracy of voltage over-limit prevention and control is highly dependent on the quality of data preprocessing and state characterization.

[0003] However, existing data processing technologies suffer from the following drawbacks: Spatiotemporal feature extraction is fragmented from the physical topology, often treating node monitoring data as independent time series or relying on static graph aggregation, making it difficult to simultaneously capture dynamic power flow coupling and load spatiotemporal distribution. This may result in an initial representation lacking a global characterization of the grid's dynamic operating ground state. Anomaly detection and repair mechanisms are crude, often relying on fixed threshold discrimination and linear interpolation, lacking fine-grained contextual attention modeling. This easily distorts the nonlinear fluctuation trajectory of source loads and erases key prior features such as meteorological and topological characteristics, potentially inducing control misjudgments. The construction of interval uncertainty sets is rigid, often setting static boundaries based on historical extreme values, failing to consider the spatial distribution differences of node voltage sensitivity and the constraint backpropagation effect under dynamic topology reconstruction of the distribution network, and lacking a dynamic bias correction mechanism, potentially causing optimization strategies to become overly conservative or lack robustness. Multimodal feature fusion suffers from dimensionality mismatch, often employing shallow splicing to construct reinforcement learning state vectors, without establishing feature alignment and dimensionality reduction mapping under a unified latent space, easily leading to the curse of dimensionality and gradient oscillations, potentially restricting the convergence efficiency of multi-agent adversarial games and meta-learning frameworks. Summary of the Invention

[0004] The technical problem to be solved by this invention is to provide a method and system for preventing and controlling voltage overruns in distribution networks based on reinforcement learning. Based on data repair, dynamic boundary construction, dual-agent game theory and meta-reinforcement learning, a policy solution set is obtained and deployed in real time to achieve closed-loop dynamic control of voltage overruns in distribution networks.

[0005] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows: Firstly, a reinforcement learning-based method for preventing and controlling voltage overruns in distribution networks, the method comprising: Multi-source heterogeneous operation data of the power distribution network are collected, and an initial spatiotemporal representation sequence is extracted using a spatiotemporal graph neural network. Based on the initial spatiotemporal representation sequence, anomalies are located through an attention mechanism and repaired by calling a radial basis function model, resulting in a time-series feature sequence containing meteorological environment, network topology, and equipment operating condition characteristics. A dynamic voltage stability boundary hypersurface is constructed based on time-series feature sequences, and spatial constraint compensation coefficients are calculated. The boundary bias is corrected for the interval uncertainty set covering source load fluctuations using the spatial constraint compensation coefficients to obtain the corrected uncertainty boundary. The meteorological environment, network topology, equipment operating conditions and the corrected uncertainty boundary are fused together to form a multidimensional state vector. A multidimensional state vector is input into a two-agent game environment. The first agent outputs a sequence of preventive control actions. The second agent searches for boundary uncertainty values ​​and Lagrange multipliers within the corrected uncertainty boundary based on the preventive control action sequence, and constructs a reward and punishment feedback. Based on the reward and punishment feedback, the two-agent strategy is iteratively updated until the game converges, and the interaction trajectory is obtained. Based on interactive trajectories and reward / punishment feedback, meta-reinforcement learning is used for multi-objective decomposition pre-training and random weight sampling to obtain a multi-objective Pareto front policy solution set. The multi-objective Pareto front strategy solution set is deployed to the real-time control system of the distribution network. Based on the current multi-dimensional state vector matching, control commands for voltage regulation equipment are issued, and closed-loop iterative control is achieved through state transition updates and voltage monitoring.

[0006] Secondly, a reinforcement learning-based distribution network voltage over-limit prevention control system includes: The anomaly repair module is used to collect multi-source heterogeneous operation data of the distribution network, extract the initial spatiotemporal representation sequence through a spatiotemporal graph neural network, locate anomalies through an attention mechanism and call the radial basis function model for repair, and obtain a time-series feature sequence containing meteorological environment, network topology and equipment operating conditions. The correction module is used to construct a dynamic voltage stability boundary hypersurface based on time-series feature sequences and calculate spatial constraint compensation coefficients; the spatial constraint compensation coefficients are used to correct the boundary bias of the interval uncertainty set covering source load fluctuations to obtain the corrected uncertainty boundary; the meteorological environment, network topology, equipment operating condition features and the corrected uncertainty boundary are fused together to form a multi-dimensional state vector. The game convergence module is used to input the multidimensional state vector into the dual-agent game environment. The first agent outputs a sequence of preventive control actions. The second agent searches for the boundary uncertainty value and Lagrange multiplier within the corrected uncertainty boundary based on the preventive control action sequence, and constructs a reward and punishment feedback. Based on the reward and punishment feedback, the dual-agent strategy is iteratively updated until the game converges, and the interaction trajectory is obtained. The optimization module is used to perform multi-objective decomposition pre-training and random weight sampling based on interaction trajectory and reward / punishment feedback, and obtain the multi-objective Pareto front policy solution set. The control module is used to deploy the multi-objective Pareto front strategy solution set to the real-time control system of the distribution network. It matches and issues control commands to the voltage regulation equipment based on the current multi-dimensional state vector, and realizes closed-loop iterative control through state transition updates and voltage monitoring.

[0007] Thirdly, a computing device includes: One or more processors; A storage device for storing one or more programs that, when executed by one or more processors, cause the one or more processors to implement the method.

[0008] Fourthly, a computer-readable storage medium storing a program that, when executed by a processor, implements the method.

[0009] The above-described solution of the present invention has at least the following beneficial effects: By collecting multi-source heterogeneous operation data from the distribution network and extracting the initial spatiotemporal representation sequence through a spatiotemporal graph neural network, the spatiotemporal coupling characteristics of the data can be fully explored. An attention mechanism is used to locate data anomalies, and a radial basis function model is invoked to complete the repair, preserving core features such as meteorological, topological, and equipment operating conditions to form a complete time-series feature sequence. This improves the integrity and usability of the data, providing a reliable data foundation for subsequent state representation. Based on the time-series feature sequence, a dynamic voltage stability boundary hypersurface is constructed, and spatial constraint compensation coefficients are calculated to correct the boundary bias of the interval uncertainty set, adapting to source-load fluctuations and dynamic changes in the power grid. Multi-dimensional features are fused with the corrected uncertainty boundary to assemble a multi-dimensional state vector, achieving multi-modal feature adaptation and improving the state vector's ability to represent the power grid's operating state. The multi-dimensional state vector is input into a dual-agent game environment. Through the collaborative operation of the two agents, the sequence of preventive control actions and the value of the search boundary uncertainty are obtained. By employing Lagrange multipliers to construct a reward and punishment feedback system that aligns with real-world operation, and iteratively updating the strategy until game convergence, the value of data interaction can be fully utilized to form a complete interaction trajectory, providing realistic sample data for subsequent meta-learning training. Based on the interaction trajectory and reward and punishment feedback, meta-reinforcement learning is used for multi-objective decomposition pre-training and random weight sampling. This efficiently utilizes the previously accumulated interaction data to achieve multi-objective optimization decomposition and weight allocation, obtaining a multi-objective Pareto front policy solution set, expanding policy diversity, and improving the policy's adaptability to different operating scenarios. The policy solution set is deployed to the real-time control system, and the optimal control command is matched and issued in conjunction with the current multi-dimensional state vector. Closed-loop iteration is achieved through state transition updates and voltage monitoring, enabling rapid policy invocation and dynamic optimization. This fully utilizes real-time operating data to improve the policy library, ensuring the continuity of data processing and control execution, and enhancing the stability and long-term adaptability of voltage control. Attached Figure Description

[0010] Figure 1 This is a flowchart illustrating the reinforcement learning-based distribution network voltage over-limit prevention and control method provided in an embodiment of the present invention.

[0011] Figure 2 This is a schematic diagram of a power distribution network voltage over-limit prevention and control system based on reinforcement learning, provided by an embodiment of the present invention. Detailed Implementation

[0012] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0013] like Figure 1 As shown, embodiments of the present invention propose a distribution network voltage over-limit prevention and control method based on reinforcement learning, the method comprising the following steps: Step 100: Collect multi-source heterogeneous operation data of the distribution network, and extract the initial spatiotemporal representation sequence through a spatiotemporal graph neural network; based on the initial spatiotemporal representation sequence, locate anomalies through an attention mechanism and call the radial basis function model for repair, to obtain a time-series feature sequence containing meteorological environment, network topology and equipment operating condition characteristics; Step 200: Construct a dynamic voltage stability boundary hypersurface based on the time-series feature sequence and calculate the spatial constraint compensation coefficient; use the spatial constraint compensation coefficient to correct the boundary bias of the interval uncertainty set covering source load fluctuations to obtain the corrected uncertainty boundary; integrate the meteorological environment, network topology, equipment operating condition features with the corrected uncertainty boundary and assemble them into a multi-dimensional state vector. Step 300: Input the multidimensional state vector into the dual-agent game environment. The first agent outputs a sequence of preventive control actions. The second agent searches for boundary uncertainty values ​​and Lagrange multipliers within the corrected uncertainty boundary based on the preventive control action sequence to construct a reward and punishment feedback. Based on the reward and punishment feedback, iteratively update the dual-agent strategy until the game converges to obtain the interaction trajectory. Step 400: Based on the interaction trajectory and reward / punishment feedback, meta-reinforcement learning is used to perform multi-objective decomposition pre-training and random weight sampling to obtain the multi-objective Pareto front policy solution set; Step 500: Deploy the solution set of the multi-objective Pareto front strategy to the real-time control system of the distribution network, match and issue control commands for voltage regulation equipment according to the current multi-dimensional state vector, and realize closed-loop iterative control through state transition update and voltage monitoring.

[0014] In this embodiment of the invention, efficient acquisition and normalization of multi-source heterogeneous operating data are achieved. Spatiotemporal representations of the data are extracted through a spatiotemporal graph neural network, preserving spatial correlations and temporal features. Anomalies are located using an attention mechanism, and noise and missing data are repaired using a radial basis function model, forming a high-quality temporal feature sequence that integrates multi-dimensional information. A dynamic voltage stability boundary is constructed based on the temporal feature sequence, and the uncertainty boundary is corrected. A multi-dimensional state vector is obtained by fusing the data, reflecting the real-time operating status and potential fluctuation risks of the distribution network. A dual-agent game environment is built to complete collaborative decision-making and strategy iteration convergence, enabling the prevention and control strategy to actively adapt to source-load fluctuations and uncertain interference, thereby improving the adaptive capability of the control strategy. Meta-reinforcement learning is used to obtain a multi-objective Pareto front strategy solution set, reducing computational consumption and ensuring the quality of the solution set. The multi-objective Pareto front strategy solution set is applied to the real-time control system to achieve closed-loop prevention and dynamic adjustment of voltage over-limit, ensuring that the operating voltage of the power grid remains within a stable range, and improving the safety, stability, and control efficiency of the distribution network.

[0015] In a preferred embodiment of the present invention, step 100 involves collecting multi-source heterogeneous operation data of the distribution network, extracting an initial spatiotemporal representation sequence using a spatiotemporal graph neural network, and based on the initial spatiotemporal representation sequence, locating anomalies through an attention mechanism and calling a radial basis function model for repair, thereby obtaining a time-series feature sequence containing meteorological environment, network topology, and equipment operating condition characteristics, including: Step 101: Collect voltage amplitude, active power, reactive power, switch status, and corresponding meteorological monitoring data of each node in the distribution network within a continuous time window, as multi-source heterogeneous operation data. Specifically, this includes: defining a continuous time acquisition window according to a preset time resolution. The preset time resolution is determined comprehensively based on the actual operation and control needs of the distribution network, the fluctuation characteristics of distributed new energy output, and the load change pattern. This time resolution is used to define the time interval scale of data sampling, which can match the real-time requirements of voltage over-limit prevention and control; synchronously traversing and collecting data from all electrical monitoring nodes in the distribution network, sequentially reading the voltage amplitude, active power, reactive power, and switch status of each node at each sampling time, and simultaneously collecting meteorological monitoring data such as wind speed, light intensity, and temperature aligned with each sampling time; performing time-series alignment and format normalization processing on the collected electrical and meteorological data to complete the unified collection and organization of multiple types of data, forming multi-source heterogeneous operation data of the distribution network.

[0016] Step 102: Construct a dynamic topology graph with electrical nodes as vertices and line impedance and real-time power flow coupling relationships as edges. Map multi-source heterogeneous operating data to the node and edge attributes of the dynamic topology graph according to time steps to obtain graph structure time-series data. Specifically, this includes: traversing all electrical nodes of the distribution network and assigning them unique identifiers, treating each numbered node as an independent topology vertex of the dynamic topology graph; retrieving the physical line connection relationships between vertices one by one, reading the impedance parameter values ​​of the corresponding lines, calculating the transmission correlation degree of real-time power flow between vertices based on line transmission characteristics, updating the connectivity status between vertices in real time according to changes in the distribution network switch status; and synchronously refreshing line impedance parameters and power flow. The correlation degree value uses the real-time changing connection relationships, impedance parameters, and power flow correlation degree as dynamic topological edges between topological vertices to establish a dynamic topological graph structure that can adaptively update in real time following switch changes, power flow transfers, and topological reconfiguration. Multi-source heterogeneous operating data is split sequentially according to the sampling time step, and the node voltage amplitude, active power, reactive power, and switch status data at each time step are assigned values ​​and mapped to the attribute dimensions of the corresponding topological vertices. The line power flow values ​​and impedance parameters are assigned values ​​and mapped to the attribute dimensions of the corresponding topological edges, completing the one-to-one correspondence binding between the data at each time step and the graph structure, thereby forming a graph structure time-series data that changes continuously with time steps.

[0017] Step 103: Input the graph-structured time-series data into a pre-trained spatiotemporal graph neural network. Aggregate spatial features of the neighborhood information of each node through graph convolutional layers to obtain an updated spatial feature tensor. Specifically, the pre-trained spatiotemporal graph neural network consists of a multi-layer graph convolutional module, a time-series feature extraction module, and a fully connected feature output module cascaded together. The graph convolutional module contains multiple graph convolutional layers, and the time-series feature extraction module contains multiple gated recurrent unit layers. This network uses historical multi-source heterogeneous operation data of the distribution network and labeled spatiotemporal features as training samples. An offline supervised learning method is employed. Multiple rounds of iterative training are conducted, progressively transferring and optimizing the internal weight parameters of the network to continuously fit the spatial correlation and temporal variation patterns of power grid nodes until the model loss function converges and stabilizes, forming a pre-trained network model with high-precision spatiotemporal feature extraction capabilities. Graph-structured temporal data is input layer by layer into the pre-trained spatiotemporal graph neural network. Spatial features are extracted from the self-attributes and neighboring node attributes of each node through graph convolutional layers. The extracted neighborhood features are then weighted, aggregated, and nonlinearly transformed to enhance the expression of electrical coupling relationships and spatial distribution characteristics between nodes. Specifically, the graph convolutional layers extract spatial features from each node... In the Spatial feature aggregation of layers Perform as follows: ; in, node In the The feature vector of the layer, For nodes The set of neighboring nodes, For nodes In the The feature vector of the layer, For the first The learnable weight matrix of the first layer , The second layer is the input feature dimension. ; For bias vectors, It is the mean aggregation function. It is the ReLU activation function. The layer index of the graph convolutional layer is used to complete the unified update and fusion of spatial features of all nodes, and finally obtain a spatial feature tensor with regular dimensions and complete information.

[0018] Step 104 involves inputting the updated spatial feature tensor into the gated recurrent unit layer in time step order, propagating and updating the hidden state along the time dimension, and outputting the initial spatiotemporal representation vector of each node step by step. Specifically, this includes: sorting the updated spatial feature tensor temporally according to the time step order and sequentially feeding it into the gated recurrent unit layer of the spatiotemporal graph neural network; the gated recurrent unit layer performs feature filtering and state initialization on the spatial features input at the current time step, combining the hidden state from the previous time step to complete information fusion, and using an internal gate control mechanism to perform retention, forgetting, and update calculations on the temporal features, progressively propagating and iteratively optimizing the hidden state along the time dimension to fully capture the changing trends and dependencies of the data within consecutive time steps; the gated recurrent unit layer outputs the initial spatiotemporal representation vector of each node step by step. The hidden state update at each time step is done using the following formula: conduct: ; ; ; ; in, For the first Vectorized representation of the spatial feature tensor input at each time step. For the best The momentary hidden state of the step, and These are updating the gate and resetting the gate vector, respectively. In the candidate hidden state, For the sigmoid function, For element-wise multiplication, and These are learnable parameters; the initial spatiotemporal representation vector output at each time step is the value of that step. , As a time step index, after the calculation is completed at each time step, the initial spatiotemporal representation vector of the corresponding node, which integrates spatial features and temporal features, is output synchronously to ensure that the vector fully carries the spatiotemporal correlation information of the power grid operation.

[0019] Step 105 involves arranging all initial spatiotemporal representation vectors in chronological order to form an initial spatiotemporal representation sequence. Specifically, this includes: matching and classifying all initial spatiotemporal representation vectors based on the sampling time step identifier to establish a one-to-one correspondence between time steps and representation vectors; sorting all representation vectors in ascending order of sampling time steps, removing duplicate vectors and filling in missing positions; performing dimension normalization and feature normalization on each vector to ensure consistent feature dimensions across all vectors; sequentially verifying temporal connection and feature continuity between adjacent vectors according to the sorting results to ensure the complete transmission relationship of vectors in the temporal dimension; and concatenating all verified representation vectors sequentially to form an initial spatiotemporal representation sequence that perfectly matches the acquisition window duration, has unified feature dimensions, and is logically coherent in temporal sequence.

[0020] Step 106: Obtain the confidence interval boundary obtained in advance based on normal historical operation data; input the initial spatiotemporal representation sequence into the anomaly detection unit based on the self-attention mechanism, and the anomaly detection unit calculates the attention weight of each time step in turn to obtain the attention weight sequence. Specifically, this includes: the confidence interval boundary obtained in advance is a numerical range determined by statistical analysis based on the historical normal operation data of the distribution network, used to distinguish between normal data and abnormal data. This boundary is obtained by retrieving historical normal operation data of the distribution network during long-term stable operation, performing batch statistics and distribution calculations on similar spatiotemporal representation vectors, analyzing the central tendency and dispersion of the data, and determining the numerical distribution range under normal operation conditions through numerical calculations; input the initial spatiotemporal representation sequence into the anomaly detection unit with a built-in self-attention mechanism, and the anomaly detection unit performs feature correlation calculations on the representation vector of each time step in the sequence, successively evaluates the correlation strength between the current vector and the overall sequence, and obtains the corresponding attention weight value; the self-attention mechanism... Force mechanism on the first Attention weights at each time step The calculation is as follows: ; ; ; in, This represents the key matrix in the self-attention mechanism. Represents the query matrix. Represents a value matrix, This is the initial spatiotemporal representation sequence matrix; For time steps, To represent the dimension of the vector, For learnable projection matrices, For the projection dimension, the value is... , This is the attention weight matrix. For the first The attention weight vectors of each time step are integrated sequentially according to their original time order to form an attention weight sequence that has the same length as the initial spatiotemporal representation sequence and corresponds one-to-one with the position.

[0021] Step 107: Compare each weight value in the attention weight sequence with the confidence interval boundary. When the weight value at any time step exceeds the confidence interval boundary, it is determined that the data corresponding to that time step is abnormal, and the time step is marked as abnormal. The comparison and marking of all time steps are completed sequentially to obtain an abnormal location mask. The time step positions in the abnormal location mask correspond to the abnormal data segments to be repaired. Specifically, this includes: traversing the attention weight sequence point by point according to the chronological order of the time steps, extracting the attention weight value corresponding to each time step in turn, and comparing the extracted weight values ​​with the predetermined confidence interval boundary item by item to determine whether the weight value is within the confidence interval. Within the normal range; when the weight value of any time step is higher than the upper limit of the confidence interval or lower than the lower limit of the confidence interval, the initial spatiotemporal representation data corresponding to that time step is immediately determined to be abnormal, and the time step is simultaneously marked as abnormal and the corresponding position is recorded; when the weight value is within the confidence interval, the data of that time step is determined to be in a normal state and the original marking is maintained; the weight comparison and state marking of all time steps are completed according to the above traversal rules, and the normal and abnormal marking results of all time steps are integrated in an orderly manner to obtain an abnormal position mask with the same length as the initial spatiotemporal representation sequence and corresponding position. The time step position marked as abnormal in the mask is the location of the abnormal data segment that needs to be repaired later.

[0022] Step 108: Based on the anomaly location mask, extract the masked anomaly data segments from the initial spatiotemporal representation sequence; for each anomaly data segment, extract feature vectors from the normal time steps before and after it to form a support point set. Specifically, this includes: based on the anomaly location mask, performing position-by-position retrieval and positioning on the initial spatiotemporal representation sequence, filtering out all time step segments marked as an anomaly state by the mask, and extracting the corresponding anomaly data segments in continuous time sequence; for each independent anomaly data segment, backtracking a specified number of normal time steps forward to extract feature vectors of the adjacent normal time period, and simultaneously extending forward a specified number of normal time steps to extract feature vectors of the subsequent stable time period; performing validity screening and normalization on the extracted forward and backward feature vectors, eliminating redundant features and unifying the vector dimensions; and combining the screened valid feature vectors in chronological order to form a support point set for supporting subsequent data repair calculations, ensuring that the support point set can fully reflect the normal operation characteristics and patterns surrounding the anomaly segment.

[0023] Step 109: Input the support point set into the radial basis function model for weighted interpolation to calculate the repair value; replace the original outlier values ​​at the corresponding positions in the initial spatiotemporal representation sequence with the repair value; after traversing all outlier segments, obtain the repaired initial spatiotemporal representation sequence as the time series feature sequence. This time series feature sequence retains the original meteorological environment, network topology, and equipment operating condition characteristics, specifically including: pre-selecting multiple continuous periods of long-term stable operation of the distribution network according to a unified time resolution, filtering out operating records with missing information, disturbances, and abnormal markings, and retaining records with stable operating conditions, normal topology, and new energy output fluctuations. Complete and undamaged time-series records with smooth dynamics are used to pre-collect qualified and normal time-series feature data. The selected compliant data serves as the base sample for initializing the radial basis function model. The basic mapping structure of the model is defined based on the sample feature distribution. Offline iterative training is conducted using batches of historical samples to continuously correct the model's internal correlation parameters and mapping rules, fully adapting to the nonlinear data change patterns under source load fluctuations and stabilizing the model's local feature fitting and interpolation performance. The support point set is input into the trained and finalized radial basis function model, which uses a Gaussian kernel function. Repair value for each abnormal point Calculate using the following formula: ; in, This is the time step index corresponding to the anomaly point. For the support point concentration A normal point in time step index. For the number of support points, The kernel width parameter is adaptively determined based on the time span of the support point set. , To normalize the weight coefficients, the model performs a comprehensive correlation analysis of normal features within the support point set, conducts multi-level weighted fitting based on the changing trends of adjacent temporal features, and calculates the specific repair values ​​for abnormal points based on the local feature mapping relationship. The calculated repair values ​​are used to replace the original abnormal values ​​at the corresponding positions in the initial spatiotemporal representation sequence, and all abnormal data segments are processed cyclically segment by segment to complete the correction and replacement of abnormal points across the entire domain in an orderly manner. Finally, a coherent and feature-complete initial spatiotemporal representation sequence is obtained and used as a temporal feature sequence. This temporal feature sequence fully retains the meteorological environment, network topology, and equipment operating condition-related feature information in the original data.

[0024] In a preferred embodiment of the present invention, step 200 involves constructing a dynamic voltage stability boundary hypersurface based on a time-series feature sequence and calculating spatial constraint compensation coefficients; using the spatial constraint compensation coefficients to correct the boundary bias of the interval uncertainty set covering source load fluctuations, thereby obtaining a corrected uncertainty boundary; and fusing meteorological environment, network topology, equipment operating condition characteristics, and the corrected uncertainty boundary to assemble a multi-dimensional state vector, including: Step 201 involves extracting the voltage amplitude and phase angle time-series data of the source-end grid-connected nodes, the grid hub nodes, and the load boundary nodes from the time-series feature sequence, and merging them to obtain the voltage dynamic response trajectory of each node. Specifically, this includes: performing a node-by-node traversal search on the time-series feature sequence to select three types of core nodes: source-end grid-connected nodes, grid hub nodes, and load boundary nodes; extracting the voltage amplitude and time-series data and phase angle time-series data of the above three types of nodes at each time step; classifying and aggregating the extracted data of the same type of nodes; merging and integrating the voltage amplitude and time-series data and phase angle time-series data of the three types of nodes according to the order of the time steps; synchronously verifying the data time-series continuity; eliminating time-series misaligned data; and finally obtaining a complete and continuous voltage dynamic response trajectory for each node, providing comprehensive node voltage dynamic data support for the subsequent construction of voltage stability boundaries.

[0025] Step 202 involves dimensionality reduction embedding of the voltage dynamic response trajectories of each node, mapping the high-dimensional trajectories to a low-dimensional manifold space to construct a multi-dimensional voltage response manifold. Specifically, this includes: performing a dimensional traversal analysis of the high-dimensional features contained in each trajectory based on the voltage dynamic response trajectories of each node; statistically analyzing the information entropy and variance of each dimension's features; selecting key features with high information entropy, stable variance, and the ability to characterize the core laws of voltage dynamic changes; and eliminating invalid features with low information entropy and high redundancy. The dimensionality requirements of the preset low-dimensional manifold space are determined, taking into account the actual operational needs of the distribution network, data processing efficiency, and the adaptability of subsequent feature fusion. The preset range is 8 to 15 dimensions, with the specific dimensions determined based on the number of distribution network nodes, the complexity of voltage response features, and the subsequent model computation requirements. During the preset process, the feature dimensions of historical operational data, existing computing power, and actual business scenarios are fully considered to ensure that the preset dimensions can both completely retain the core information of voltage dynamic changes and effectively avoid the problem of redundancy in high-dimensional data processing.

[0026] A dimensionality reduction embedding method is employed to compress the dimensionality of the high-dimensional voltage dynamic response trajectory after preserving key features. Following a preset dimensionality requirement for the low-dimensional manifold space, the high-dimensional trajectory is progressively mapped to this low-dimensional manifold space. During the mapping process, the feature retention rate is progressively verified, and the changes in key features before and after mapping are compared dimensionally. If the feature deviation exceeds a preset range, the mapping parameters are adjusted promptly for remapping, ensuring that the core information of the voltage dynamic changes is not lost. The dimensionality reduction embedding uses a manifold learning algorithm based on an adaptive kernel function, specifically: constructing an affinity matrix based on the Euclidean distance between each time step. ,in , To achieve adaptive kernel width, the first value of each sample point is taken. The median of nearest neighbor distances. ;right Row normalization yields the Markov transition matrix. ;right Perform eigenvalue decomposition, taking the 2nd to the 3rd eigenvalues. There are eigenvectors, where... In this embodiment, the dimension of the low-dimensional manifold space is preset. As a result of dimensionality reduction embedding, through the above series of operations, a multidimensional voltage response manifold that can reflect the dynamic correlation of voltages at each node is finally constructed, effectively solving the problem of redundancy in high-dimensional data processing in existing technologies.

[0027] Step 203: Apply a preset voltage stability threshold as a constraint boundary to the multidimensional voltage response manifold to obtain a constrained voltage response manifold; extract a hypersurface from the constrained voltage response manifold to obtain a dynamic voltage stability boundary hypersurface. Specifically, this includes: setting a voltage stability threshold that meets the actual operating requirements of the distribution network based on the multidimensional voltage response manifold. This voltage stability threshold refers to a reasonable range of values ​​within which the voltage amplitude of each node in the distribution network can remain stable and not cause voltage exceedance or instability risks during normal dynamic operation; the preset process is carried out in conjunction with the distribution network voltage level, line parameters, equipment rated operating conditions, and power grid safety operation procedures. Historical normal operating voltage data of the corresponding voltage level of the distribution network is retrieved, the conventional distribution range of voltage amplitude of each node is statistically analyzed, and the threshold benchmark range is determined after removing abnormal fluctuation data in conjunction with the power grid safety and stability operation standards; the benchmark range is dynamically fine-tuned according to the new energy penetration rate, load fluctuation characteristics, and topology complexity of the distribution network, and finally, a voltage stability threshold suitable for the current power grid operating state is determined.

[0028] The voltage stability threshold is applied as a constraint to the multidimensional voltage response manifold. The manifold regions exceeding the threshold range are initially screened to obtain a constrained voltage response manifold. All local hyperpolygon patches generated across the simplex are stitched together according to the adjacency relationship set during simplex subdivision: adjacent simplexes sharing the same spatial interface have consistent interpolation intersections on the common surface. These patches are seamlessly connected to obtain a complete, piecewise linear dynamic voltage stability boundary hypersurface. This dynamic voltage stability boundary hypersurface is the boundary between voltage stability and instability in the constrained voltage response manifold, realizing the dynamic characterization of the voltage stability boundary.

[0029] Step 204 involves calculating the Gaussian curvature and average curvature of the dynamic voltage stability boundary hypersurface to obtain the curvature tensor distribution. Specifically, this includes determining a preset sampling interval for the dynamic voltage stability boundary hypersurface. This sampling interval is preset based on the voltage response characteristics of the distribution network, the complexity of the hypersurface, and the accuracy requirements of subsequent curvature calculations. During the preset process, the spatial distribution range of the hypersurface, the coverage efficiency of discrete sampling points, and the operability of subsequent curvature calculations are comprehensively considered, along with the distribution density of actual operating data of the distribution network. Typically, the preset sampling interval is controlled between 0.05 and 0.1 voltage units to adapt to the spatial scale and characteristic distribution of the dynamic voltage stability boundary hypersurface.

[0030] The dynamic voltage stability boundary hypersurface is discretized globally. A uniform grid is then created on the hypersurface according to a preset sampling interval. The spatial coordinates of each sampling point are determined and marked, resulting in several uniformly distributed discrete sampling points. Simultaneously, the distribution density of the sampling points is verified to ensure complete coverage of the entire hypersurface and avoid calculation errors caused by sparse sampling in local areas. For each discrete sampling point, Gaussian curvature and average curvature are calculated point-by-point. The spatial coordinates and voltage characteristic data of adjacent sampling points are obtained. Curvature values ​​are calculated based on the characteristic correlations between adjacent points. The Gaussian curvature and average curvature values ​​of each sampling point are calculated and recorded point-by-point. The compliance of the calculation results is verified concurrently during the calculation process. If the curvature value of a sampling point deviates from the reasonable range of surrounding sampling points, the data of that sampling point and its adjacent points are retrieved and recalculated to ensure the accuracy of the curvature calculation results.

[0031] The curvature data of all sampling points are collected in an orderly manner according to their corresponding spatial coordinates. Based on the spatial distribution law of the dynamic voltage stability boundary hypersurface, the curvature data is classified and organized to sort out the spatial distribution trend of the curvature data and form a complete and orderly curvature tensor distribution. Through the above curvature quantification analysis process, the curvature differences in different regions of the dynamic voltage stability boundary hypersurface are captured, providing a basis for the subsequent division of stability sensitive regions.

[0032] Step 205: Based on the locations of local extrema in the curvature tensor distribution, perform gradient curvature segmentation on the dynamic voltage stability boundary hypersurface to obtain several sub-regions with different stability sensitivities. Specifically, this includes: based on the curvature tensor distribution, performing a full-domain point-by-point traversal search of the distribution data, comparing the value of each curvature data with its surrounding adjacent curvature data, and determining whether the curvature data is greater than or less than all its adjacent data, thereby locating all local extrema in the curvature tensor distribution; simultaneously recording the spatial coordinates corresponding to each local extrema, validating the located extrema, eliminating false extrema caused by data fluctuations, and ensuring the accuracy of local extrema location; using the verified local extrema as the segmentation benchmark, determining each extremum... The gradient curvature change direction corresponding to the value point is used to perform gradient curvature segmentation on the hypersurface of the dynamic voltage stability boundary. During the segmentation process, the change amplitude of the curvature gradient is monitored in real time, and the segmentation accuracy is dynamically adjusted according to the curvature gradient change trend. The segmentation accuracy is appropriately increased in areas with drastic curvature gradient changes and appropriately decreased in areas with gentle curvature gradient changes, ensuring that the curvature distribution in each sub-region tends to be uniform, thereby maintaining a consistent stability sensitivity in each sub-region. After segmentation, the spatial range of each sub-region is defined and marked, and the spatial positional relationship and stability sensitivity level of each sub-region are sorted out. Finally, several sub-regions with different stability sensitivities are obtained, effectively realizing the refined partitioning management of the stability boundary and providing a partitioning basis for subsequent sub-region constraint analysis.

[0033] Step 206: For each sub-region, calculate the node voltage sensitivity matrix based on the power flow Jacobian matrix of each node within the sub-region. Using the node voltage sensitivity matrix as weights, perform constraint propagation based on topology reconstruction, projecting the voltage constraints of the sub-region boundary nodes back along the power flow path to the source-end grid-connected nodes to obtain the constraint inversion components of the sub-region. Summate the constraint inversion components of all sub-regions to obtain the spatial constraint compensation coefficient. Specifically, for the current sub-region, based on the real-time topology and power flow state of the distribution network, calculate the power flow Jacobian matrix of the sub-region using Newton-Raphson power flow in polar coordinates. For those containing For a subregion with nodes, the matrix form of the linearized power flow correction equation is: ; Among them, ΔP=[ΔP1,ΔP2,…,ΔP n ]ᵀ represents the active power imbalance vector at each node; ΔQ=[ΔQ1,ΔQ2,…,ΔQ n ]ᵀ represents the reactive power imbalance vector at each node; Δθ=[Δθ1,Δθ2,…,Δθ n]ᵀ represents the node voltage phase angle correction vector; ΔV=[ΔV1,ΔV2,…,ΔV n ]ᵀ represents the node voltage magnitude correction vector; H=∂ΔP÷∂θ, N=∂ΔP÷∂V, M=∂ΔQ÷∂θ, L=∂ΔQ÷∂V represent the Jacobian matrix. The four sub-matrices; by eliminating the phase angle correction Δθ from equation (1), the partial derivative matrix of the node voltage magnitude with respect to the node injected reactive power is derived, i.e., the voltage sensitivity matrix S: ; Where, S∈ℝ n ˣ n Matrix elements Represents a node When the injected unit reactive power changes, the node The change in voltage amplitude, as accurately characterized by this sensitivity matrix, reflects the electrical coupling strength of node voltages within the sub-region; therefore, it is directly used as the weight matrix for subsequent constraint propagation. Weight allocation can utilize diagonal elements. Characterizing the node's own sensitivity, or the elements after row normalization. As the influence weights between nodes, the voltage sensitivity matrix is ​​obtained. Then, using it as the weight matrix, the back projection algorithm is executed. The core of the algorithm is: to backtrack the voltage constraints of the sub-region boundary nodes along the power flow path of the distribution network segment by segment, and to use the voltage constraints of each node as a weight matrix. The provided weighted correction constraint bias is ultimately projected onto the source-end grid-connected node; based on the actual topology and power flow direction of the sub-region, the reverse power flow node sequence from the boundary node to the source-end grid-connected node is identified, for example: boundary node - segmented switch node - tie switch node - source-end grid-connected node; the voltage operation constraint of the boundary node is then applied, such as... As initial constraint signal ,in The voltage amplitude at the boundary node; along the reverse path, the constraint signal is transferred from the node. Passed to node ,in, For the node index in the reverse power flow node sequence, during transmission, the voltage coupling weight between the two nodes is calculated based on the sensitivity matrix S. Alternatively, other coefficients allocated based on actual sensitivity contributions can be used to correct the constraint value using the following formula: ; in, To the current downstream node Given the voltage constraint value, Assigned to upstream nodes after this weighted adjustment The voltage constraint value is determined, and if there are sensitivity changes due to topology reconstruction in the path, the local sensitivity parameters are updated in real time, and the weights are recalculated. When the constraint signal is projected onto the source-end grid-connected node, the constraint inversion component of that sub-region is obtained. ,in, ; Taking a 10kV feeder sub-region as an example: the power flow direction is source node - tie switch - sectional switch - boundary node; the reverse projection path is boundary node - sectional switch - tie switch - source node. Assuming the boundary node voltage constraint is 0.95~1.05pu, the weight is obtained from the corresponding elements of the S matrix, and the weight is adjusted segment by segment. Finally, the inversion component is obtained at the source node. Traverse all sub-regions, calculate their respective constraint inversion components, and perform algebraic summation on all components. Perform compliance verification on the summation result. If an abnormal deviation occurs in a component of a certain sub-region, backtrack and recalculate the sensitivity or projection process until all are reasonable. Finally, the global spatial constraint compensation coefficient is obtained.

[0034] Step 207: Obtain the interval uncertainty set pre-statistically calculated based on historical source-load fluctuation data. The interval uncertainty set includes the upper and lower bounds of the new energy power generation and load power at each time step. Using spatial constraint compensation coefficients, nonlinear bias corrections are applied to the upper and lower bounds of the interval uncertainty set, with the correction magnitude proportional to the compensation coefficient, to obtain the corrected uncertainty boundary. Specifically, this includes: determining the specific statistical process for the pre-statistically calculated interval uncertainty set, which is pre-statistically calculated based on long-term historical source-load fluctuation data of the distribution network. During the statistical analysis, a historical data statistical period is defined, and all source-load fluctuation data under normal operating conditions of the distribution network within that period are retrieved, and missing or abnormal data are filtered out. During periods of disturbance and abnormal operating conditions, source and load data are retained, ensuring that the power generation and load data of new energy sources are complete and their fluctuations conform to normal patterns. The selected valid historical data are then segmented by time step, and batch statistics are performed for each time step. The power generation and load data of all historical moments under that time step are sorted numerically to determine the maximum and minimum values ​​of power generation and load power under each time step. The maximum values ​​of the two types of power corresponding to each time step are used as the upper bound of the interval, and the minimum values ​​of the two types of power are used as the lower bound of the interval. All upper and lower bound data are integrated in chronological order of time steps to complete the pre-statistics of the interval uncertainty set. The pre-calculated interval uncertainty set is retrieved to ensure that it fully includes the upper and lower bounds of the new energy power generation and load power at each time step. Spatial constraint compensation coefficients are then used to perform nonlinear bias corrections on the upper and lower bounds of the interval uncertainty set. During the correction process, the correction magnitude is adjusted according to the size of the compensation coefficients to ensure that the correction magnitude is proportional to the compensation coefficients. Specifically, the nonlinear bias correction is performed using the following formula: ; ; in, and The first The original upper and lower bounds of the uncertainty set of the time step interval. , These are the corrected upper and lower bounds; The value of the spatial constraint compensation coefficient at this time step; This represents the standard deviation of the historical source load fluctuation data at this time step. For time step index, To correct the strength coefficient, this embodiment takes... =0.5; As the scaling factor for the compensation coefficient, this embodiment takes... =2.0, which is determined based on the matching relationship between the nonlinear saturation characteristics of the hyperbolic tangent function and the physical boundary of the electrical coupling strength of the distribution network; The function ensures that the correction amount does not exceed the standard deviation range, thus achieving nonlinear boundary compression. After the correction is completed, the validity of the corrected boundary is verified, boundary values ​​that exceed the reasonable operating range are removed, and the boundary is corrected again to finally obtain a corrected uncertainty boundary that fits the dynamic operating conditions of the distribution network.

[0035] Step 208 involves aligning the meteorological environment, network topology, and equipment operating condition features retained in the time-series feature sequence with the corrected uncertainty boundary to obtain a comprehensive feature set. Specifically, this includes: extracting the retained meteorological environment, network topology, and equipment operating condition feature data from the time-series feature sequence based on the time-series feature sequence and the corrected uncertainty boundary; performing a full domain traversal on each of the three types of feature data to identify the specific feature items and data dimensions included in each type of feature data; statistically analyzing the number of dimensions, numerical distribution range, and data format of each type of feature data; and determining the dimensional differences and adaptation difficulties of each type of feature data. The feature dimensions of the corrected uncertainty boundary are then regularized, extracting all feature items included in the corrected uncertainty boundary, and analyzing their number of dimensions and data format. A set of feature dimensions with moderate quantity and suitable for subsequent feature fusion needs is selected as the benchmark dimension standard. All three types of feature data and the feature dimensions for correcting uncertainty boundaries are adjusted to this benchmark dimension standard. During the adjustment process, feature data with insufficient dimensions are supplemented to match the actual operating rules of the corresponding features. Feature data with redundant dimensions is simplified and invalid redundant dimensions are eliminated to ensure that the dimensions of all types of feature data are completely consistent after adjustment, effectively eliminating dimensional barriers between different types of features. After completing the dimensional alignment of multiple types of features, the validity of the aligned feature data is verified. Numerical deviations and data misalignments that occur during the dimensional alignment process are investigated and corrected in a timely manner. Finally, a comprehensive feature set containing all types of features with unified dimensions is obtained, laying a solid foundation for subsequent feature fusion.

[0036] Step 209 involves performing a feature concatenation operation on the comprehensive feature set to fuse it into a comprehensive feature matrix. Specifically, this includes: determining a preset order and method for sorting the features in the comprehensive feature set. This preset order is set based on the actual operating needs of the distribution network, the importance of the features, and the convenience of subsequent processing. During the preset process, it is fully considered that meteorological environmental characteristics are a fundamental influencing factor on power grid operation, network topology and equipment operating conditions are the core manifestations of the power grid's operating status, and correcting uncertainty boundary features is the key to adapting to dynamic operation. Therefore, the preset order is determined to be meteorological environment, network topology, equipment operating conditions, and correcting uncertainty boundary. The preset order is combined with the daily operating rules of the distribution network, references historical feature processing experience, and takes into account the correlation of features and the efficiency of subsequent processing to ensure that the preset order fits the actual application scenario.

[0037] The four categories of features within the comprehensive feature set—meteorological environment, network topology, equipment condition, and corrected uncertainty boundary—are categorized and sorted. The specific data composition and temporal attributes of each feature category are determined, and the features are sorted according to a pre-defined order. During the sorting process, the temporal sequence and spatial correlation logic of each feature are followed to ensure complete aggregation of similar features and orderly sorting of different categories. The compliance of the sorting results is verified to avoid feature misalignment or category confusion. For each sorted feature category, a feature concatenation operation is performed. The meteorological environment feature data, which is ranked first, is retrieved and arranged according to the chronological order of time steps. The network topology feature data is then concatenated segment by segment after the meteorological environment feature data according to the corresponding time steps. Finally, the equipment condition features and corrected uncertainty boundary features are sequentially concatenated and integrated into the feature sequence according to the corresponding time step relationships. During the splicing process, the temporal step correspondence and spatial association attributes of each type of feature are checked in real time to ensure that the temporal association between different types of features is not interrupted and the spatial association is not destroyed. After each type of feature is spliced, the integrity of the current spliced ​​feature sequence is checked to check for problems such as missing features, splicing misalignment, and data redundancy. If problems are found, the splicing order and splicing method are adjusted in time to supplement missing features, remove redundant data, and correct misalignment problems. After all types of features are spliced, the overall spliced ​​feature sequence is checked in its entirety to confirm that the features are complete, the temporal sequence is coherent, and the spatial association is reasonable. Finally, a comprehensive feature matrix with a regular structure and complete features is formed. By unifying the data organization form, the processing difficulty of subsequent vector transformation and model operation is reduced, and regularized data support is provided for the construction of subsequent multi-dimensional state vectors.

[0038] Step 210 involves expanding the comprehensive feature matrix into a one-dimensional feature vector by row, and performing batch normalization on the one-dimensional feature vector to obtain the multi-dimensional state vector required for reinforcement learning. Specifically, this includes: expanding the comprehensive feature matrix row by row, starting from the first row of the matrix, extracting all feature data for each row, and sequentially connecting the features within each row to the end of the expanded feature sequence of the previous row according to the original arrangement order of the features, gradually transforming it into a one-dimensional feature vector; during the expansion process, the arrangement order and connection position of the features in each row are checked in real time, preserving the original correlation relationship and time step correspondence of each feature, and verifying the continuity of feature connection segment by segment to avoid problems such as feature misalignment, order reversal, or data omission, ensuring that the expanded one-dimensional feature vector completely replicates all feature information of the comprehensive feature matrix.

[0039] The one-dimensional feature vector is batch normalized by iterating through and extracting all feature values, statistically analyzing the numerical distribution range of all feature values, determining the target numerical range for normalization, and adjusting the value of each feature value one by one. Based on the correspondence between the original feature values ​​and the target range, the magnitude of each feature value is gradually adjusted until all feature values ​​are uniformly standardized to the same numerical range. During the adjustment process, the adjustment magnitude of each feature value is simultaneously calculated to ensure that the adjusted feature values ​​meet the target range requirements without losing the core differences of the original features, thus mitigating computational interference caused by differences in feature scales. After normalization, the normalized one-dimensional feature vector undergoes a global compliance check, verifying that the value of each feature value is within the target range and identifying any numerical anomalies, over-adjustments, or disruptions to feature relationships. If anomalies are found, the normalization parameters of the corresponding feature values ​​are readjusted and recalculated until all feature values ​​meet the requirements. This process ultimately yields a multi-dimensional state vector that meets the needs of reinforcement learning, standardizing the feature data scale, ensuring the stable operation of subsequent multi-agent game and meta-learning frameworks, and providing regular state input support for the optimization of subsequent control strategies.

[0040] In a preferred embodiment of the present invention, step 300 involves inputting a multidimensional state vector into a dual-agent game environment, with the first agent outputting a sequence of preventive control actions; the second agent, based on the preventive control action sequence, searching for boundary uncertainty values ​​and Lagrange multipliers within the corrected uncertainty boundary to construct a reward-penalty feedback; and iteratively updating the dual-agent strategy based on the reward-penalty feedback until the game converges, yielding the interaction trajectory, including: Step 301 involves inputting the multidimensional state vector into the first agent in the two-agent game environment to obtain a sequence of preventive control actions. Specifically, this includes: performing a full traversal of the multidimensional state vector, checking each feature data in the vector one by one, and confirming that the data is complete and without missing or misaligned elements before inputting it completely into the first agent in the two-agent game environment. The first agent employs a deep neural network structure, consisting of an input layer, a hidden layer, and an output layer. The input layer receives the multidimensional state vector, the hidden layer performs multi-level correlation mining and feature transformation on the input features, and the output layer outputs the sequence of preventive control actions. The training method combines historical operating data of the distribution network with simulated operating scenarios. Historical state data and corresponding optimal control action data under different operating conditions of the distribution network are retrieved to construct a training dataset. Historical state data is used as input, and corresponding optimal control actions are used as labels to iteratively train the first agent. During training, network parameters are continuously adjusted until the deviation between the control actions output by the agent and the labeled actions reaches a preset requirement, thus completing the training.

[0041] The first agent's built-in policy network has an input dimension consistent with the multi-dimensional state vector obtained earlier, ensuring that the input data can be fully adapted to the policy network. The output dimension matches the number of distribution network preventive control actions, enabling the synchronous output of multiple control actions. The first agent sorts through the input multi-dimensional state vector, distinguishes the feature information of different dimensions, and combines it with the control logic of the actual operation of the distribution network. It calls its own policy network to perform correlation analysis on relevant features such as weather, topology, and equipment operating conditions in the vector, and filters out the core features related to voltage over-limit prevention. Based on these core features, and according to the power flow transmission law and control requirements of the distribution network, it gradually obtains the corresponding preventive control actions, and finally integrates them to form a continuous and standardized preventive control action sequence. This sequence is ensured to be adaptable to the current operating state of the power grid, providing a reliable action basis for the subsequent search work of the second agent.

[0042] Step 302: Input the preventive control action sequence and the corrected uncertainty boundary into the second agent. Search within the corrected uncertainty boundary for the boundary uncertainty value that maximizes the risk of voltage exceeding the limit, and obtain the corresponding Lagrange multiplier. Specifically, this includes: retrieving the preventive control action sequence and simultaneously acquiring the corrected uncertainty boundary data, aligning the two dimensions to ensure consistent data format and time synchronization; inputting the preventive control action sequence and the corrected uncertainty boundary into the second agent, which adopts a deep neural network structure adapted to the first agent, consisting of an input layer, a hidden layer, and an output layer. The input layer receives the two types of data: the preventive control action sequence and the corrected uncertainty boundary. The hidden layer mines the correlation between the two types of data and performs feature transformation. The output layer outputs the boundary uncertainty value and the corresponding Lagrange multiplier. The training method combines historical operating data and simulated operating conditions of the distribution network. It collects Lagrange multipliers of boundary uncertainties under different operating scenarios of the distribution network, as well as corresponding control effect data, to construct a training dataset. The preventive control action sequence and the corrected uncertainty boundary are used as input samples, and the corresponding optimal boundary uncertainty value and Lagrange multiplier are used as labels. The second agent is iteratively trained, and the network parameters are continuously adjusted until the agent can select the value that maximizes the risk of voltage exceedance and the corresponding Lagrange multiplier, thus completing the training. The policy network built into the second agent has the same input dimension as the feature dimension of the preventive control action sequence and the corrected uncertainty boundary, ensuring that the input data can be fully adapted to the network. The output dimension corresponds to two types of data: boundary uncertainty value and Lagrange multiplier, achieving synchronous output.

[0043] The second intelligent agent performs a full-domain scan of the input data to determine the range and constraints of the corrected uncertainty boundary. Based on these constraints, it searches point by point within the boundary range for the boundary uncertainty value that maximizes the risk of voltage exceeding the limit. It verifies the voltage operating state corresponding to each candidate value to determine whether it will lead to voltage exceeding the limit. Simultaneously, it calculates the Lagrange multiplier corresponding to each candidate value and, combined with the power balance characteristics and topology of the distribution network, selects the Lagrange multipliers that meet the requirements. Finally, it determines the boundary uncertainty value and the corresponding Lagrange multiplier, providing data support for the subsequent construction of reward and punishment feedback.

[0044] Step 303: Based on the prevention and control action sequence, boundary uncertainty values, and Lagrange multipliers, and combined with the power balance constraints of the distribution network, construct a reward and punishment feedback system. This specifically includes: summarizing the prevention and control action sequence, boundary uncertainty values, and corresponding Lagrange multipliers; systematically reviewing and verifying the three types of data one by one to determine the specific content and time sequence of each type of data, ensuring that there are no missing or misaligned data; classifying and organizing the data according to data type, archiving the prevention and control action sequence in time step order, sorting the boundary uncertainty values ​​by size, and classifying the Lagrange multipliers according to their correspondence, ensuring the temporal consistency and dimensional compatibility of all types of data, and avoiding data confusion and incorrect correspondence; and combining the power balance constraints of the distribution network... The requirements are as follows: The matching degree between the sequence of preventive and control actions and the boundary uncertainty values ​​must be verified one by one. Based on the previously determined voltage sensitivity weights, weights must be assigned to each control action, each boundary value, and each Lagrange multiplier. The influence weights corresponding to various types of data must be determined, incorporating the execution effect of preventive and control actions, the fluctuation range of boundary uncertainty values, and the adaptability of Lagrange multipliers into the calculation scope, and the correlation of each data point must be verified one by one. Simultaneously, in conjunction with the actual power transmission situation of the distribution network, the balance between power input and output must be verified, and any power imbalances or abnormal values ​​must be investigated. If data deviations are found, relevant parameters must be adjusted promptly to ensure that the calculation results conform to actual operational needs.

[0045] Based on this, the calculation results of various data are integrated, and combined with the operating rules and control requirements of the distribution network, a reward and punishment feedback system that can truly reflect the control effect and operating status is constructed, and the specific standards for rewards and punishments are determined. Depending on the magnitude, the reward and punishment feedback is constructed according to the following reward and punishment function: ; in, In order to prevent and control actions With boundary uncertainty value Voltage deviation of a single node; The cost of adjusting the control action; This refers to the violation of power balance constraints in the distribution network. These are Lagrange multiplier vectors; , The weighting coefficients ensure that the reward and punishment feedback corresponds to each control action and operating status, without omitting key data or causing redundant calculations. This provides reliable data support for subsequent agent strategy iteration and interaction trajectory analysis, ensuring the orderly progress of subsequent game and optimization work.

[0046] Step 304: Input the reward and punishment feedback into the policy network of the first and second agents, and calculate the original gradient based on the reward and punishment feedback; construct a safe gradient flow constraint boundary based on the original gradient, specifically including: inputting the complete reward and punishment feedback into the policy network of the first and second agents, standardizing the reward and punishment feedback data, adjusting the values ​​of the feedback data one by one to standardize them to the numerical range suitable for the policy network, ensuring that the data can adapt to the input requirements of the policy network, and avoiding the impact of data scale differences on the accuracy of subsequent calculations; based on the specific values ​​of the reward and punishment feedback, calculating the gradient of each relevant parameter in the policy network one by one, capturing the correlation between each policy parameter and the reward and punishment feedback, verifying the degree of influence of parameter changes on the reward and punishment feedback, and thus obtaining the complete original gradient data, while verifying the compliance of the original gradient and investigating abnormal gradient fluctuations.

[0047] After obtaining the original gradient, the target data is compared point-by-point with the convex set boundary using the convex set projection algorithm. The data is adjusted to fit inside the convex set, thus correcting the gradient data constraints and adapting to the constraints of dynamic operation of the distribution network. Combining the constraints of dynamic operation of the distribution network, the convex set projection algorithm is used to analyze the patterns and ranges of gradient changes, determining the convex set constraint range for gradient changes under dynamic operation of the distribution network. This range is used as the preset convex set for the convex set projection algorithm. Here, the preset convex set range refers to the gradient fluctuation interval set based on the actual operating constraints of the distribution network, voltage safety thresholds, and gradient calculation requirements. It typically ranges from -0.8 to 0.8, but can be flexibly adjusted according to the distribution network voltage level, load fluctuation amplitude, and control accuracy requirements. The method involves combining historical gradient data of the distribution network, the safe operating range of voltage, and previously determined voltage sensitivity parameters to analyze the intervals of gradient changes, eliminate the ranges corresponding to abnormal gradient values, and simultaneously consider the constraints of dynamic operation of the distribution network, balancing the stability of strategy updates and control accuracy. Finally, a preset convex set range is determined to ensure that this range can cover normal gradient fluctuations while effectively constraining abnormal gradients, preventing gradient disorder from affecting strategy updates. The original gradient data is input point by point into the convex set projection algorithm, and each gradient data point is checked to determine whether it falls within the preset convex set range. Gradient data exceeding the convex set range is projected and adjusted, mapping it to the convex set boundary or interior, ensuring that the gradient data conforms to the dynamic operation constraints of the distribution network.

[0048] Taking the gradient calculation of the strategy network corresponding to a 10kV distribution network as an example, the preset convex set constraint range for gradient changes is -0.8 to 0.8. The original gradient data is input point by point into the convex set projection algorithm. If a certain original gradient value is 0.9, which exceeds the preset convex set range, the gradient value is adjusted to 0.8 through the convex set projection algorithm. If a certain original gradient value is -0.9, it is adjusted to -0.8. Gradient data within the range remain unchanged. Through this process, the reasonable constraint boundary of the safe gradient flow is determined, the allowable range of gradient changes is determined, and finally the safe gradient flow constraint boundary is constructed, providing a standard and basis for subsequent gradient correction, ensuring that the subsequent gradient correction work is accurate and efficient, and meets the actual operation needs of the distribution network.

[0049] Step 305 involves projecting and correcting the original gradient using the secure gradient flow constraint boundary to obtain the secure projected gradient; then updating the network parameters using the secure projected gradient to obtain the secure adversarial strategy kernel. Specifically, this includes: retrieving the original gradient and the secure gradient flow constraint boundary; performing a comprehensive analysis of both types of data to determine the specific numerical distribution and temporal correlation of the original gradient; simultaneously verifying the value range of the secure gradient flow constraint boundary to ensure that the dimensions of the two types of data are consistent and the temporal sequence is synchronized, avoiding subsequent comparison calculations due to data misalignment or dimension mismatch; comparing the original gradient with the secure gradient flow constraint boundary point by point, extracting the specific value of each original gradient, and verifying it against the upper and lower limits of the constraint boundary to determine whether each original gradient value exceeds the reasonable range specified by the constraint boundary; synchronously recording original gradients that exceed the range and those that do not, classifying and archiving them to ensure no omissions or misjudgments.

[0050] For original gradients that exceed the constraint range, projection correction is performed according to the requirements of the safe gradient flow constraint boundary. Combining the core logic of the convex set projection algorithm, the gradient values ​​are gradually adjusted according to the magnitude of the gradient exceeding the boundary, correcting the excess gradient values ​​to within the constraint boundary range. The corrected gradient values ​​are verified point by point to ensure that the corrected gradients not only meet the constraint requirements but also retain the core characteristics of the original gradients, thus obtaining a single safe projected gradient. For original gradients that do not exceed the constraint range, their original values ​​are directly retained and included in the safe projected gradient set one by one. At the same time, the included gradient data is checked to avoid data duplication, misalignment, and other problems, and finally integrated to form a complete safe projected gradient set.

[0051] The secure projection gradient is used to update the policy network parameters of the first and second agents parameter by parameter. This determines the current value and update priority of each parameter in the policy network. Based on the actual needs of the dynamic operation of the distribution network and referring to the previously determined voltage sensitivity parameters and power balance constraints, each policy network parameter is adjusted one by one. The parameter values ​​are adjusted according to the magnitude of the secure projection gradient to ensure that the adjustment direction aligns with the network operation patterns. During parameter adjustment, the adaptability of each updated parameter is simultaneously verified to determine whether the updated parameters are suitable for the current meteorological environment, network topology, and equipment operating conditions of the distribution network. If a mismatch is found between the updated parameters and the actual operating state of the network, the parameter values ​​are fine-tuned until all parameters are adapted to the current network operating state. After all parameters are updated, a full-domain verification of the updated policy network is performed to identify any anomalies that occurred during the parameter update process, ensuring stable operation of the policy network and reasonable parameter settings. This ultimately forms a secure adversarial policy core, providing stable and reliable policy support for subsequent iterative optimization and ensuring the orderly progress and convergence efficiency of the dual-agent game.

[0052] Step 306: Calculate the current reward / penalty feedback value based on the security countermeasure strategy kernel. If the convergence condition is not met, return to the step of inputting the multidimensional state vector into the first agent to output the prevention and control action sequence, and continue iterating; otherwise, stop iterating. Concatenate the prevention and control action sequence, boundary uncertainty values, and Lagrange multipliers output from each iteration according to time steps to obtain the interaction trajectory. Specifically, this includes: calling the security countermeasure strategy kernel, checking and verifying the parameters of the security countermeasure strategy kernel one by one to confirm that the kernel parameters are normal and without missing values; substituting each parameter of the security countermeasure strategy kernel into the reward / penalty feedback calculation, combined with the power balance constraints and electrical... The system calculates the sensitivity parameters and corrects the uncertainty boundary. It retrieves the prevention and control action sequence, boundary uncertainty values, and Lagrange multiplier data for each time step, and calculates the specific reward and punishment feedback values ​​for each time step. The calculation results for each time step are recorded synchronously to ensure that the calculation process is traceable and the calculation results are accurate. The system compares and analyzes the currently calculated reward and punishment feedback values ​​one by one to determine whether the reward and punishment feedback values ​​for each time step tend to be stable and whether there are no more obvious fluctuations. Then it determines whether the two-agent game meets the convergence condition. The judgment results are verified point by point during the comparison process to avoid misjudgment or omission.

[0053] If the convergence condition is not met, the current calculation process is immediately stopped, and the process returns to step 301. The multidimensional state vector is then fully input into the first agent, which re-organizes the feature data and obtains a new sequence of preventive control actions. The entire calculation process from step 302 to step 305 is repeated sequentially, and operations such as boundary uncertainty value search, Lagrange multiplier calculation, reward and punishment feedback construction, gradient calculation and correction, and policy network parameter update are carried out again to start the next iteration. During the iteration process, the number of iterations and the calculation data of each stage are recorded simultaneously for subsequent traceability and verification.

[0054] If the convergence condition is met, the iteration process is stopped immediately. The output of each iteration, including the sequence of preventative control actions, boundary uncertainty values, and Lagrange multipliers, is retrieved and categorized according to the iteration number and time step sequence. Each type of data is then time-sequentially checked to ensure a one-to-one correspondence between the three types of data in each iteration, avoiding data misalignment or time sequence confusion. After categorization, the data is gradually spliced ​​together according to the time step sequence. During splicing, the continuity and correspondence of the data are checked in real time to ensure that the time sequence correlation of each type of data is not disrupted. After splicing, a full-domain verification is performed on the overall data to check for missing data, splicing misalignment, redundancy, etc. If any anomalies are found, timely adjustments and corrections are made to ultimately form a complete and standardized interaction trajectory. This provides reliable data support for subsequent distribution network voltage over-limit prevention and control, ensuring the orderly implementation of subsequent control strategy optimization work.

[0055] In a preferred embodiment of the present invention, step 400 above, based on the interaction trajectory and reward / punishment feedback, employs meta-reinforcement learning for multi-objective decomposition pre-training and random weight sampling to obtain a multi-objective Pareto front policy solution set, including: Step 401 involves acquiring the interaction trajectory and reward / penalty feedback as a training sample set for meta-reinforcement learning. Specifically, this includes: sorting out the interaction trajectory, extracting the preventive control actions, boundary uncertainty values, and relevant parameter data contained within it, and summarizing the changing patterns of specific values ​​in the reward / penalty feedback. The two types of data are then categorized and organized according to time steps to ensure a one-to-one correspondence between the interaction trajectory data and the reward / penalty feedback data for each time step, with no missing data or incorrect correspondence. After organization, all data is filtered to remove abnormal and invalid data, retaining only valid data that truly reflects the operating status and control effect of the distribution network, forming a complete training sample set. This provides data support for subsequent meta-reinforcement learning training and ensures that the sample set covers data under different operating conditions of the distribution network, improving the comprehensiveness of the training.

[0056] Step 402: Based on the training sample set, a multi-objective decomposition method is used to randomly sample multiple weight vectors on the unit simplex; each weight vector is normalized to obtain a normalized weight vector; using the normalized weight vectors, the multiple optimization objectives of voltage over-limit prevention control are decomposed into multiple single-objective sub-problems, resulting in a set of sub-problems and the corresponding weight combinations for each sub-problem. Specifically, this includes: sorting out the training sample set, extracting all optimization objectives related to voltage over-limit prevention control contained in the sample set, determining the core requirements and data characteristics of each optimization objective, and verifying the data integrity and temporal consistency in the sample set to ensure that the sample data can provide reliable support for multi-objective decomposition; and using the multi-objective decomposition method to carry out various computational processing tasks in an orderly manner. Specifically, the multi-objective decomposition method adopts the MOEA / D framework based on Chebyshev aggregation: for the first... Normalized weight vectors in, To optimize the target number, the value is selected. =3, decomposing the multi-objective problem into scalar optimization subproblems: ; in, This is the policy parameter vector; For the first One objective function value; The vector of the ideal reference point; For the first The weight vector of the th weight vector One component; The core idea of ​​the Chebyshev aggregation method, which is the decomposed scalar optimization function, is: for each normalized weight vector... Calculate all Weighted bias of each objective Take the maximum value as the scalar optimization function value. By minimizing This forces the optimization process to focus on the objective that deviates furthest from the ideal point, achieving balanced optimization across multiple objectives. To obtain a Pareto front solution set covering different optimization preferences, multiple differentiated normalized weight vectors need to be sampled on the unit simplex. A random sampling method is used to generate candidate weight vectors. Each sampled weight vector is compared with the already sampled set of weight vectors to determine if their numerical differences meet a preset standard. The preset standard is the minimum Euclidean distance threshold between weight vectors, with a value of [value missing]. ,in To optimize the target number, in this embodiment... That is, when the Euclidean distance between the newly sampled weight vector and any vector in the set of already sampled weight vectors is less than 1. If a duplicate is detected, resampling is performed. This threshold ensures that... Approximately 100 unique weight vectors are uniformly covered on the unit simplex. After sampling, all generated weight vectors are classified and archived, and the sampling number and numerical characteristics of each weight vector are marked one by one.

[0057] After sampling, each weight vector is normalized. The sum of the values ​​of each weight vector is calculated one by one. Each value in each weight vector is divided by the sum of the values ​​of that vector. The range of values ​​of each weight vector is adjusted point by point so that the sum of the values ​​of all weight vectors tends to be consistent and on the same order of magnitude. This effectively eliminates the calculation interference caused by the difference in the dimensions between different weight vectors. After the normalization process of each weight vector is completed, the compliance of its normalized value is verified to avoid problems such as numerical anomalies and incomplete normalization.

[0058] By combining multiple optimization objectives of voltage over-limit prevention and control, the core parameters and constraints of each optimization objective are systematically analyzed. The normalized weight vector is matched with each optimization objective. Based on the numerical distribution of the weight vector, the original multi-objective optimization problem is gradually decomposed into multiple independent single-objective sub-problems. Each single-objective sub-problem corresponds to a core dimension of the original multi-objective optimization. During the decomposition process, the optimization direction, core requirements, and data range of each sub-problem are determined one by one. At the same time, based on the normalized weight vector, the weight combination corresponding to each sub-problem is determined, and the degree of influence of each weight in the sub-problem is determined. After the decomposition is completed, all single-objective sub-problems are reviewed and verified to ensure that each sub-problem corresponds to a certain optimization dimension of the original multi-objective, without omissions or repetitions. Finally, they are integrated to form a complete set of sub-problems, providing a clear training direction and data basis for subsequent meta-model training.

[0059] Step 403: Initialize the meta-model parameters based on the training sample set and the sub-problem set; update the meta-model parameters through gradient iteration to obtain the trained meta-model. Specifically, this includes: constructing a meta-model based on the training sample set and the sub-problem set. The meta-model construction needs to be combined with the voltage over-limit prevention and control requirements of the distribution network, and adopts a deep neural network architecture, consisting of an input layer, a hidden layer, and an output layer. The input layer is used to receive the interaction trajectory and reward / penalty feedback data from the training sample set, the hidden layer is used to mine the correlation features between the sample data and the sub-problems and realize feature transformation, and the output layer is used to output the strategy parameters adapted to each sub-problem. During the construction process, the number of nodes, activation methods, and connection relationships of each network layer are determined in combination with the operation rules of the distribution network to ensure that the meta-model structure can adapt to the training requirements of the sub-problems after multi-objective decomposition, while taking into account the model's computational efficiency and adaptation accuracy, thus completing the construction of the meta-model.

[0060] Initialize the meta-model parameters, determine the basic parameter settings of the meta-model, and determine the initial values ​​of the parameters based on the operation law and control requirements of the distribution network. Ensure that the initial parameters conform to the dynamic operation characteristics of the distribution network and avoid excessive deviation of the initial parameters that may cause subsequent training failure. The training process of the meta-model is carried out simultaneously with the implementation process. The training is based on the training sample set and the sub-problem set as the core basis, and meta-training is performed through gradient iteration. Specifically, the training sample set is split into each sub-problem in the sub-problem set, the weight combination of the corresponding sub-problem is matched with the sample data, and input into the constructed meta-model. In each iteration, the optimization requirements of the sub-problems and the reward / penalty feedback numbers are considered. According to the process, the parameters of the meta-model are adjusted one by one to gradually optimize the model's adaptability to the operating conditions of the distribution network. During the iteration process, the rationality of the parameter adjustments is continuously verified, and the matching degree between the adjusted parameters and the operating constraints of the distribution network and the optimization objectives of the sub-problems is checked one by one to avoid parameter deviations causing model failure. At the same time, the parameter adjustment range and model output results of each iteration are recorded to facilitate subsequent traceability and optimization. After multiple gradient iterations, the iteration stops when the strategy parameters output by the meta-model can stably adapt to all sub-problems and the deviation between the model output results and the sample data reaches the preset requirements. The trained meta-model is then obtained, ensuring that the meta-model can respond to different operating conditions and control requirements.

[0061] Step 404 involves sampling multiple target weight vectors at fixed intervals within the target weight space, and fine-tuning the trained meta-model using a finite number of iterations to obtain the fine-tuned strategy model corresponding to each target weight vector. Specifically, this includes: determining the value range and distribution pattern of the target weight space; determining the sampling interval and number of samples based on the actual needs of distribution network voltage control; performing uniform sampling at fixed intervals, selecting a weight value at each fixed interval to obtain multiple different target weight vectors; ensuring significant differences in the values ​​of each weight vector during sampling to avoid overly concentrated sampling leading to incomplete coverage; recording the specific values ​​and corresponding optimization preferences of each target weight vector to ensure that the sampled weight vectors cover different optimization preference scenarios and meet the needs of different operating conditions; and after sampling, reviewing and verifying all target weight vectors, removing weight vectors with abnormal or duplicate values ​​to ensure that each target weight vector corresponds to different optimization preferences, providing a reliable basis for subsequent model fine-tuning.

[0062] For each sampled target weight vector, it is matched one by one with the trained meta-model. The initial parameters and structure of the meta-model are retrieved to determine the adjustable parameter range and adjustment rules of the model. Combined with the distribution network operation constraints and the optimization preferences of the corresponding weight vector, the direction and magnitude of parameter fine-tuning are determined. During the fine-tuning process, the number of iteration steps is controlled, and only a limited number of iterations are performed without changing the core structure and adaptation logic of the meta-model. The focus is on adjusting the model's output strategy to adapt to the optimization requirements of the corresponding weight vector. During the adjustment process, the compliance of parameter adjustment is verified in real time to ensure that the fine-tuned model parameters meet the distribution network operation constraints and are consistent with the optimization preferences of the corresponding weight vector.

[0063] For each target weight vector whose model fine-tuning is completed, all parameters of the fine-tuned strategy model are recorded in detail, including the adjusted model structure, parameter values, and appropriate optimization preferences. The completeness and accuracy of the recorded information are checked one by one to avoid parameter omissions or recording errors. After the fine-tuning work for all target weight vectors is completed, all fine-tuned strategy models are reviewed to check the adaptability and parameter compliance of each model, ensuring that each fine-tuned strategy model can match the optimization requirements of the corresponding weight vector, thus guaranteeing the diversity and adaptability of the models.

[0064] Step 405: Deploy each fine-tuned strategy model to the distribution network simulation environment and collect the cumulative optimization target value of each model to obtain a target value vector set; perform Pareto dominance comparison on the target value vector set, extract all non-dominated solutions, and aggregate them into a multi-objective Pareto front strategy solution set. Specifically, this includes: deploying all fine-tuned strategy models to the distribution network simulation environment to simulate the actual operation scenarios of the distribution network, including the operation status under different load levels and weather conditions, recording various data in real time during the operation of each fine-tuned strategy model, focusing on the cumulative optimization target value, integrating the operation data of all models to form a target value vector set, and ensuring that the operation data of each model is complete and traceable.

[0065] The convex hull algorithm is introduced to assist in policy selection. The convex hull algorithm refers to constructing the smallest convex polygon containing all target value vectors to define the distribution range of all target values, clearly presenting the differences in optimization effects corresponding to different policies. At the same time, the distribution range of the optimal policy can be determined by the convex hull boundary, providing an intuitive basis for subsequent policy selection. The specific implementation process is as follows: all target value vectors are imported into the convex hull algorithm to determine the coordinate position of each target value vector. With each target value vector as a vertex, the smallest convex polygon containing all vertices, i.e., the convex hull, is constructed. Through the convex hull, the differences in optimization effects of different policies can be clearly distinguished, invalid policies can be eliminated, and the policy combination with the best comprehensive performance can be selected.

[0066] Assume there are three fine-tuned strategy models, each corresponding to a target value vector. The first vector represents excellent load control and moderate voltage stability; the second vector represents excellent voltage stability and moderate load control; and the third vector represents moderate load control and voltage stability. These three vectors are imported into a convex hull algorithm to construct a convex hull containing three vertices. The vectors covered by the convex hull boundary are the non-dominated solutions, which are the target value vectors corresponding to the strategy with the best overall performance. After completing the convex hull analysis, the target value vector set is compared group by group, performing Pareto dominance comparisons to filter out all non-dominated solutions—solutions where no other solution is superior to this one across all optimization objectives. All non-dominated solutions are then compiled and sorted according to the priority of the optimization objectives, ultimately forming a multi-objective Pareto front strategy solution set. This ensures that the solution set can cover the optimal control strategies under different operating scenarios, providing reliable strategy support for subsequent voltage exceedance prevention control in the distribution network. Furthermore, the application of the convex hull algorithm further improves the accuracy and efficiency of strategy selection, avoiding the selection of invalid strategies and ensuring that the obtained strategy solution set can adapt to the needs of different operating conditions.

[0067] In a preferred embodiment of the present invention, step 500 involves deploying the multi-objective Pareto front strategy solution set to the real-time control system of the distribution network, matching and issuing control commands for voltage regulation equipment based on the current multi-dimensional state vector, and achieving closed-loop iterative control through state transition updates and voltage monitoring, including: Step 501 involves deploying each strategy model from the multi-objective Pareto front strategy solution set to the strategy library of the distribution network real-time control system. This includes: reviewing all strategy models in the multi-objective Pareto front strategy solution set, verifying the parameter settings, applicable scenarios, and control logic of each strategy model to ensure that each model is free of anomalies and duplications, and can cover the control requirements under different operating conditions of the distribution network; deploying these strategy models one by one to the strategy library of the distribution network real-time control system, ensuring compatibility with the hardware interface and data transmission protocol of the control system during deployment; debugging the calling permissions and data interaction capabilities of each strategy model to ensure that each strategy model can be correctly recognized and called by the control system; and classifying and organizing the strategy library, archiving the strategy models according to optimization direction and applicable operating conditions to facilitate rapid matching and calling in the future, laying the foundation for subsequent real-time control.

[0068] Step 502: Real-time acquisition of distribution network operation data, extraction via spatiotemporal graph neural network representation, anomaly localization via attention mechanism, and nonlinear repair of radial basis function to obtain a time-series feature sequence; construction of a dynamic voltage stability boundary hypersurface based on the time-series feature sequence, calculation of spatial constraint compensation coefficients, boundary bias correction of the interval uncertainty set to obtain a corrected uncertainty boundary, and fusion of the multidimensional features parsed from the time-series feature sequence with the corrected uncertainty boundary to obtain the current multidimensional state vector. Specifically, this includes: starting the real-time data acquisition module of the distribution network to continuously acquire operation data of each node of the distribution network, including node voltage, load power, meteorological parameters, and other relevant data. During the data collection process, the frequency and accuracy of data acquisition are controlled to avoid data loss, misalignment, or anomalies. After acquisition, the raw data is initially screened to remove obviously abnormal data points. The screened raw data is then characterized and extracted using a spatiotemporal graph neural network to uncover the spatiotemporal coupling features and intrinsic correlations in the data, avoiding the isolation of node data and the severing of physical topological connections. An attention mechanism is used to locate anomalies in the extracted feature data, identifying abnormal fluctuations and outliers in the data. Radial basis functions are used for nonlinear repair to restore the true fluctuation characteristics of the data, avoiding feature distortion caused by linear filling, and finally obtaining a complete and accurate time-series feature sequence.

[0069] Based on this, a dynamic voltage stability boundary hypersurface is constructed by combining the voltage change patterns and load fluctuation characteristics in the time-series feature sequences to accurately define the range of voltage stability operation. At the same time, spatial constraint compensation coefficients are calculated, and the boundary bias is corrected for the original interval uncertainty set by combining the dynamic topology reconfiguration characteristics of the distribution network and the differences in node voltage sensitivity, resulting in a corrected uncertainty boundary adapted to the current operating state. The multidimensional features obtained from the time-series feature sequence analysis are deeply integrated with the corrected uncertainty boundary, and the correlation information of the data from various dimensions is integrated to obtain a multidimensional state vector that can comprehensively reflect the current operating state of the distribution network, providing reliable data support for subsequent strategy matching.

[0070] Step 503: Substitute the current multidimensional state vector into the strategy library, calculate the comprehensive cost function value corresponding to each strategy model, and select the strategy model with the smallest comprehensive cost function value as the current final strategy. Specifically, this includes: retrieving the current multidimensional state vector, checking and verifying each feature data in the vector one by one to confirm that there are no missing, misaligned, or abnormal data, ensuring that the multidimensional state vector can completely and accurately reflect the real-time operating status of the current distribution network; completely importing the multidimensional state vector into the strategy library of the distribution network real-time control system, verifying the interface compatibility between the vector and the strategy library during the import process, ensuring smooth data transmission and consistent format, and avoiding the impact of data compatibility issues on subsequent matching calculations; after the import is completed, start the strategy library call command, call all deployed strategy models in the library one by one, check the parameters of each called strategy model, determine the weight allocation details, control logic, and adaptation conditions of each model, ensure that each strategy model can participate in the matching calculation normally, and record the call order and core parameters of each strategy model for subsequent traceability and verification.

[0071] For each invoked strategy model, the model's weight allocation details and control logic are matched with the current multi-dimensional state vector to determine the correlation between the features of each dimension in the vector and the model weights and control logic. Combined with the actual operating constraints of the distribution network, including power balance constraints, voltage safety thresholds, and equipment operating limits, the calculation dimensions and accounting standards of the comprehensive cost function are determined, balancing control effectiveness and operational efficiency, and avoiding cost calculation deviations caused by single-dimensional accounting. During the calculation process, each feature data in the multi-dimensional state vector is decomposed one by one, and the impact of each feature data on the comprehensive cost is calculated based on the model's corresponding weight allocation. The calculation results of each dimension are integrated to obtain the comprehensive cost function value of the strategy model under the current multi-dimensional state vector. After each model's cost calculation is completed, the calculation results are verified to check the rationality of the accounting process and the accuracy of the calculated values. If a calculation deviation is found, it is recalculated in a timely manner to ensure that the comprehensive cost function value of each strategy model is accurate. The definition is as follows: ; in, The voltage exceedance risk value is calculated as the weighted sum of squares of the voltage deviations from the stable range at each node. To control costs, the weighted sum of the squares of the adjustments made by each device is taken. This is due to network loss deviation; These are weighting coefficients extracted from the policy model. This is the current multidimensional state vector. For the control action output by the strategy, select the one that makes The smallest strategy model is selected as the final strategy. After the comprehensive cost function values ​​of all strategy models are calculated, all calculation results are summarized and sorted by numerical value. During the sorting process, the strategy model corresponding to each cost function value is determined to avoid errors in the correspondence between numerical values ​​and models. At the same time, it is checked whether there are any cases of the same cost function value. If so, the strategy model is further compared and screened in combination with the adaptation conditions and control efficiency. The strategy model with the smallest comprehensive cost function value is selected by sorting. The model's weight allocation, control logic and adaptability to the current multi-dimensional state vector are checked to confirm that the model can adapt to the current operating state and control requirements of the distribution network. This model is officially determined as the final strategy under the current operating state, ensuring that the selected strategy can optimally adapt to the current operating requirements of the distribution network, improve the pertinence of control, and provide a reliable strategy basis for the generation and issuance of subsequent control commands.

[0072] Step 504: Based on the current final strategy, obtain the control command sequence for voltage regulation equipment and send it to the corresponding voltage regulation equipment for execution. Specifically, this includes: based on the current final strategy, combined with the operating status of each node in the distribution network and the operating parameters of the voltage regulation equipment, gradually parsing out the corresponding control commands for the voltage regulation equipment, determining the adjustment direction, adjustment range, and adjustment sequence of each voltage regulation equipment, and integrating them to form a complete control command sequence; after the command sequence is generated, verify the feasibility of each command one by one to ensure that the commands comply with the operating constraints of the distribution network and the operating requirements of the equipment, and avoid command conflicts or invalid adjustments; after verification, through the real-time control system of the distribution network, accurately send the control command sequence to the corresponding voltage regulation equipment, and record the command sending time, equipment number, and command content to ensure that each device can accurately receive and execute the corresponding command, and ensure the orderly implementation of control actions.

[0073] Step 505 involves monitoring the voltage amplitude changes at each node after execution and calculating the state transition feedback signal. Specifically, this includes: during the execution of the control command, activating the distribution network voltage monitoring module to collect voltage amplitude data from each node in real time, continuously tracking voltage change trends, and recording the voltage value at each time point; after the control command is executed, comparing the voltage amplitudes of each node before and after execution, calculating key indicators such as the change in voltage amplitude and stability, and further calculating the state transition feedback signal based on these indicators. The feedback signal must accurately reflect the execution effect of the control command, clearly present the changes after voltage regulation, and provide real and reliable feedback data for subsequent strategy optimization and iteration, ensuring the continuity of closed-loop control.

[0074] Step 506: Combine the current multidimensional state vector, voltage regulation equipment control command sequence, state transition feedback signal, and next-moment state vector into an empirical sample, and store it in the buffer for periodic updating of strategy library parameters to achieve closed-loop iterative control. Specifically, this includes: organizing and verifying the current multidimensional state vector, voltage regulation equipment control command sequence, and state transition feedback signal, checking the completeness and accuracy of each type of data one by one, confirming no missing data, no corresponding errors, and no numerical anomalies, and then collecting the operating status data of the distribution network at the next moment, sorting out the operating parameters at that moment, and forming the next-moment state vector; combining these four types of data to construct a complete empirical sample, checking the correlation and accuracy of the sample data one by one to avoid missing data or corresponding errors; after verification, storing the empirical sample in the buffer of the control system; and using the empirical sample accumulated in the buffer according to a preset period to update the strategy library parameters. The parameters of the strategy library are updated and optimized to continuously adjust the adaptability of the strategy model, realize closed-loop iterative control, and ensure that the strategy library can continuously adapt to the dynamic operation changes of the distribution network, thereby improving the stability and reliability of voltage control. The preset period refers to the fixed time interval used to periodically update the strategy library parameters and realize closed-loop iteration. Its duration needs to be determined comprehensively based on the operating fluctuation frequency of the distribution network, data update requirements, and system computing power. The preset method is as follows: combine the dynamic change pattern of the daily operation of the distribution network, refer to the update frequency of historical operating data, and take into account the timeliness of strategy optimization and system operating efficiency to determine the period duration. This avoids the system computing load being too large and resources being wasted due to the period being too short, and also avoids the strategy being unable to adapt to the dynamic changes of the distribution network in a timely manner due to the period being too long. Typically, the preset period can be set to once a day, once every half day, or once an hour. The specific duration can be flexibly adjusted according to the actual operating scenario.

[0075] like Figure 2 As shown, embodiments of the present invention also provide a distribution network voltage over-limit prevention and control system based on reinforcement learning, comprising: The anomaly repair module is used to collect multi-source heterogeneous operation data of the distribution network, extract the initial spatiotemporal representation sequence through a spatiotemporal graph neural network, locate anomalies through an attention mechanism and call the radial basis function model for repair, and obtain a time-series feature sequence containing meteorological environment, network topology and equipment operating conditions. The correction module is used to construct a dynamic voltage stability boundary hypersurface based on time-series feature sequences and calculate spatial constraint compensation coefficients; the spatial constraint compensation coefficients are used to correct the boundary bias of the interval uncertainty set covering source load fluctuations to obtain the corrected uncertainty boundary; the meteorological environment, network topology, equipment operating condition features and the corrected uncertainty boundary are fused together to form a multi-dimensional state vector. The game convergence module is used to input the multidimensional state vector into the dual-agent game environment. The first agent outputs a sequence of preventive control actions. The second agent searches for the boundary uncertainty value and Lagrange multiplier within the corrected uncertainty boundary based on the preventive control action sequence, and constructs a reward and punishment feedback. Based on the reward and punishment feedback, the dual-agent strategy is iteratively updated until the game converges, and the interaction trajectory is obtained. The optimization module is used to perform multi-objective decomposition pre-training and random weight sampling based on interaction trajectory and reward / punishment feedback, and obtain the multi-objective Pareto front policy solution set. The control module is used to deploy the multi-objective Pareto front strategy solution set to the real-time control system of the distribution network. It matches and issues control commands to the voltage regulation equipment based on the current multi-dimensional state vector, and realizes closed-loop iterative control through state transition updates and voltage monitoring.

[0076] It should be noted that this system is a system corresponding to the above method. All implementation methods in the above method embodiments are applicable to this embodiment and can achieve the same technical effect.

[0077] Embodiments of the present invention also provide a computing device, including: a processor and a memory storing a computer program, wherein the computer program, when executed by the processor, performs the method described above. All implementations in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0078] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for preventing and controlling voltage overruns in distribution networks based on reinforcement learning, characterized in that, The method includes: Multi-source heterogeneous operation data of the power distribution network are collected, and an initial spatiotemporal representation sequence is extracted using a spatiotemporal graph neural network. Based on the initial spatiotemporal representation sequence, anomalies are located through an attention mechanism and repaired by calling a radial basis function model, resulting in a time-series feature sequence containing meteorological environment, network topology, and equipment operating condition characteristics. A dynamic voltage stability boundary hypersurface is constructed based on time-series feature sequences, and spatial constraint compensation coefficients are calculated. The boundary bias is corrected for the interval uncertainty set covering source load fluctuations using the spatial constraint compensation coefficients to obtain the corrected uncertainty boundary. The meteorological environment, network topology, equipment operating conditions and the corrected uncertainty boundary are fused together to form a multidimensional state vector. A multidimensional state vector is input into a two-agent game environment. The first agent outputs a sequence of preventive control actions. The second agent searches for boundary uncertainty values ​​and Lagrange multipliers within the corrected uncertainty boundary based on the preventive control action sequence, and constructs a reward and punishment feedback. Based on the reward and punishment feedback, the two-agent strategy is iteratively updated until the game converges, and the interaction trajectory is obtained. Based on interactive trajectories and reward / punishment feedback, meta-reinforcement learning is used for multi-objective decomposition pre-training and random weight sampling to obtain a multi-objective Pareto front policy solution set. The multi-objective Pareto front strategy solution set is deployed to the real-time control system of the distribution network. Based on the current multi-dimensional state vector matching, control commands for voltage regulation equipment are issued, and closed-loop iterative control is achieved through state transition updates and voltage monitoring.

2. The distribution network voltage over-limit prevention and control method based on reinforcement learning according to claim 1, characterized in that, Collect multi-source heterogeneous operation data of the distribution network, and extract the initial spatiotemporal representation sequence through a spatiotemporal graph neural network, including: Collect voltage amplitude, active power, reactive power, switch status, and meteorological monitoring data of each node in the distribution network within a continuous time window, as multi-source heterogeneous operation data; A dynamic topology graph is constructed with electrical nodes as vertices and the coupling relationship between line impedance and real-time power flow as edges. Multi-source heterogeneous operating data is mapped to the node and edge attributes of the dynamic topology graph according to time steps to obtain graph structure time series data. The graph structure time series data is input into a pre-trained spatiotemporal graph neural network, and spatial features are aggregated from the neighborhood information of each node through graph convolutional layers to obtain an updated spatial feature tensor. The updated spatial feature tensor is input into the gated recurrent unit layer in time step order, propagated along the time dimension and the hidden state is updated, and the initial spatiotemporal representation vector of each node is output step by step. Arrange all initial spatiotemporal representation vectors in chronological order to form an initial spatiotemporal representation sequence.

3. The distribution network voltage over-limit prevention and control method based on reinforcement learning according to claim 2, characterized in that, Based on the initial spatiotemporal representation sequence, anomalies are located through an attention mechanism and repaired by invoking a radial basis function model, resulting in a time-series feature sequence containing meteorological environment, network topology, and equipment operating condition characteristics, including: Obtain the confidence interval boundary based on pre-statistical analysis of normal historical operating data; input the initial spatiotemporal representation sequence into the anomaly detection unit based on the self-attention mechanism, and have the anomaly detection unit calculate the attention weights for each time step in sequence to obtain the attention weight sequence; Each weight value in the attention weight sequence is compared with the confidence interval boundary. When the weight value of any time step exceeds the confidence interval boundary, it is determined that the data corresponding to that time step is abnormal and the time step is marked as abnormal. The comparison and marking of all time steps are completed in sequence to obtain the abnormal location mask. The time step position in the abnormal state in the abnormal location mask corresponds to the abnormal data segment to be repaired. Based on the abnormal location mask, the abnormal data segments marked by the mask are extracted from the initial spatiotemporal representation sequence; for each abnormal data segment, feature vectors are extracted from the normal time steps before and after it to form a support point set; The support point set is input into the radial basis function model for weighted interpolation to calculate the repair value; the repair value is used to replace the original outlier value at the corresponding position in the initial spatiotemporal representation sequence. After traversing all outlier segments, the repaired initial spatiotemporal representation sequence is obtained as the time series feature sequence, which retains the original meteorological environment, network topology and equipment operating condition features.

4. The distribution network voltage over-limit prevention and control method based on reinforcement learning according to claim 3, characterized in that, A dynamic voltage-stability boundary hypersurface is constructed based on time-series characteristic sequences, and spatial constraint compensation coefficients are calculated, including: The voltage amplitude and phase angle time series data of the source-end grid-connected nodes, the grid hub nodes, and the load boundary nodes are extracted from the time series feature sequence, and then merged to obtain the voltage dynamic response trajectory of each node. The voltage dynamic response trajectories of each node are dimensionality-reduced and embedded, mapping the high-dimensional trajectories to the low-dimensional manifold space to construct a multidimensional voltage response manifold. A preset voltage stability threshold is applied as a constraint boundary to the multidimensional voltage response manifold to obtain a constrained voltage response manifold; a hypersurface is extracted from the constrained voltage response manifold to obtain a dynamic voltage stability boundary hypersurface. Gaussian curvature and mean curvature of the dynamic voltage stability boundary hypersurface are calculated to obtain the curvature tensor distribution. Based on the location of local extrema in the curvature tensor distribution, gradient curvature segmentation is performed on the dynamic voltage stability boundary hypersurface to obtain several sub-regions with different stability sensitivities. For each sub-region, the node voltage sensitivity matrix is ​​calculated based on the power flow Jacobian matrix of each node in the sub-region. Using the node voltage sensitivity matrix as weights, constraint propagation based on topology reconstruction is performed. The voltage constraints of the sub-region boundary nodes are back-projected along the power flow path to the source-end grid-connected nodes to obtain the constraint inversion components of the sub-region. The constraint inversion components of all sub-regions are summed to obtain the spatial constraint compensation coefficient.

5. The distribution network voltage over-limit prevention and control method based on reinforcement learning according to claim 4, characterized in that, By using spatial constraint compensation coefficients to correct the boundary bias of the interval uncertainty set covering source load fluctuations, the corrected uncertainty boundary is obtained. The meteorological environment, network topology, equipment operating conditions, and corrected uncertainty boundaries are fused to form a multi-dimensional state vector, including: A set of interval uncertainties is obtained based on historical source-load fluctuation data. The set of interval uncertainties includes the upper and lower bounds of the new energy power generation and load power at each time step. Using spatial constraint compensation coefficients, nonlinear bias corrections are made to the upper and lower bounds of the set of interval uncertainties. The correction magnitude is proportional to the compensation coefficient, thus obtaining the corrected uncertainty boundary. The meteorological environment, network topology, and equipment operating condition features retained in the time-series feature sequence are aligned with the corrected uncertainty boundary to obtain a comprehensive feature set; Perform a feature concatenation operation on the comprehensive feature set to fuse it into a comprehensive feature matrix; The comprehensive feature matrix is ​​expanded into a one-dimensional feature vector by rows, and the one-dimensional feature vector is batch normalized to obtain the multi-dimensional state vector required for reinforcement learning.

6. The distribution network voltage over-limit prevention and control method based on reinforcement learning according to claim 5, characterized in that, Input a multidimensional state vector into a two-agent game environment, and the first agent outputs a sequence of preventive control actions. The second agent, based on the sequence of preventive control actions, searches for boundary uncertainty values ​​and Lagrange multipliers within the corrected uncertainty boundary to construct a reward and punishment feedback mechanism, including: Inputting the multidimensional state vector into the first agent in a two-agent game environment yields a sequence of preventive control actions. The preventive control action sequence and the corrected uncertainty boundary are input into the second agent. Within the corrected uncertainty boundary, the boundary uncertainty value that maximizes the risk of voltage exceeding the limit is searched, and the corresponding Lagrange multiplier is obtained. Based on the sequence of prevention and control actions, the value of boundary uncertainty, and the Lagrange multiplier, a reward and punishment feedback mechanism is constructed in conjunction with the power balance constraints of the distribution network.

7. The distribution network voltage over-limit prevention and control method based on reinforcement learning according to claim 6, characterized in that, Based on reward and punishment feedback, the dual-agent strategy is iteratively updated until the game converges, yielding the interaction trajectory, including: The reward and punishment feedback is input into the policy network of the first and second agents, and the original gradient is calculated based on the reward and punishment feedback; based on the original gradient, a safe gradient flow constraint boundary is constructed. The original gradient is projected and corrected using the safety gradient flow constraint boundary to obtain the safety projected gradient; the safety projected gradient is then used to update the policy network parameters to obtain the safety adversarial policy kernel. The current reward and punishment feedback value is calculated based on the security confrontation strategy kernel. If the convergence condition is not met, the process returns to the step of inputting the multidimensional state vector into the first agent and outputting the prevention and control action sequence, and continues to iterate. Otherwise, the iteration stops. The prevention and control action sequence, boundary uncertainty value and Lagrange multiplier output by each iteration are concatenated according to the time step to obtain the interaction trajectory.

8. The distribution network voltage over-limit prevention and control method based on reinforcement learning according to claim 7, characterized in that, Based on interaction trajectories and reward / punishment feedback, meta-reinforcement learning is used for multi-objective decomposition pre-training and random weight sampling to obtain a multi-objective Pareto front policy solution set, including: The interaction trajectory and reward / punishment feedback are obtained and used as the training sample set for meta-reinforcement learning; Based on the training sample set, a multi-objective decomposition method is adopted to randomly sample multiple weight vectors on the unit simplex; each weight vector is normalized to obtain a normalized weight vector; using the normalized weight vector, the multiple optimization objectives of voltage over-limit prevention and control are decomposed into multiple single-objective sub-problems to obtain a set of sub-problems and the weight combination corresponding to each sub-problem. Based on the training sample set and the set of sub-problems, the meta-model parameters are initialized; the meta-model parameters are updated by performing meta-training through gradient iteration to obtain the trained meta-model. Multiple target weight vectors are obtained by sampling at fixed intervals in the target weight space. The trained meta-model is then fine-tuned by a finite number of iterations to obtain the fine-tuned policy model corresponding to each target weight vector. Each fine-tuned strategy model is deployed to the distribution network simulation environment for operation. The cumulative optimization target value of each model is collected to obtain a target value vector set. Pareto dominance comparison is performed on the target value vector set to extract all non-dominated solutions and aggregate them into a multi-objective Pareto front strategy solution set.

9. The distribution network voltage over-limit prevention and control method based on reinforcement learning according to claim 8, characterized in that, The multi-objective Pareto front strategy solution set is deployed to the real-time control system of the distribution network. Based on the current multi-dimensional state vector matching, control commands for voltage regulation equipment are issued. Closed-loop iterative control is achieved through state transition updates and voltage monitoring, including: Deploy the strategy models from the multi-objective Pareto front strategy solution set into the strategy library of the distribution network real-time control system. Real-time data collection of power distribution network operation is performed, and time-series feature sequences are obtained through spatiotemporal graph neural network representation extraction, attention mechanism anomaly localization, and radial basis function nonlinear repair. Based on the time-series feature sequences, a dynamic voltage stability boundary hypersurface is constructed, and spatial constraint compensation coefficients are calculated. Boundary bias correction is performed on the interval uncertainty set to obtain the corrected uncertainty boundary. The multidimensional features parsed from the time-series feature sequences are fused with the corrected uncertainty boundary to obtain the current multidimensional state vector. Substitute the current multidimensional state vector into the policy library, calculate the comprehensive cost function value corresponding to each policy model, and select the policy model with the smallest comprehensive cost function value as the current final policy. Based on the current final strategy, a sequence of control commands for voltage regulation equipment is obtained and sent to the corresponding voltage regulation equipment for execution. Monitor the voltage amplitude changes of each node after execution and calculate the state transition feedback signal; The current multidimensional state vector, the voltage regulation device control command sequence, the state transition feedback signal, and the state vector at the next moment are combined into an empirical sample and stored in a buffer for periodically updating the strategy library parameters, thereby realizing closed-loop iterative control.

10. A distribution network voltage over-limit prevention and control system based on reinforcement learning, wherein the system implements the method as described in any one of claims 1 to 9, characterized in that, include: The anomaly repair module is used to collect multi-source heterogeneous operation data of the distribution network and extract the initial spatiotemporal representation sequence through a spatiotemporal graph neural network. Based on the initial spatiotemporal representation sequence, anomalies are located through an attention mechanism and repaired by calling the radial basis function model, resulting in a time-series feature sequence containing meteorological environment, network topology, and equipment operating condition features; The correction module is used to construct a dynamic voltage stability boundary hypersurface based on the time-series feature sequence and calculate the spatial constraint compensation coefficient; the spatial constraint compensation coefficient is used to correct the boundary bias of the interval uncertainty set covering the source load fluctuation to obtain the corrected uncertainty boundary. The meteorological environment, network topology, equipment operating conditions and modified uncertainty boundary are fused together to form a multi-dimensional state vector. The game convergence module is used to input a multi-dimensional state vector into a two-agent game environment. The first agent outputs a sequence of preventive control actions. The second agent searches for boundary uncertainty values ​​and Lagrange multipliers within the corrected uncertainty boundary based on the preventive control action sequence, and constructs a reward and punishment feedback. Based on reward and punishment feedback, the dual-agent strategy is iteratively updated until the game converges, and the interaction trajectory is obtained. The optimization module is used to perform multi-objective decomposition pre-training and random weight sampling based on interaction trajectory and reward / punishment feedback, and obtain the multi-objective Pareto front policy solution set. The control module is used to deploy the multi-objective Pareto front strategy solution set to the real-time control system of the distribution network. It matches and issues control commands to the voltage regulation equipment based on the current multi-dimensional state vector, and realizes closed-loop iterative control through state transition updates and voltage monitoring.

Images