Power grid dispatch decision-making method and large model based on graph neural network

Abstract

The present invention discloses a grid dispatch decision method and a large model based on a graph neural network, which includes collecting an original grid data set, generating a fused feature tensor through adaptive distribution mapping and recursive decomposition; importing the fused feature tensor and the grid topology matrix, generating a dynamic graph embedding matrix through spectral decomposition and matrix manifold analysis; performing Markov jump process analysis and optimal transmission fusion on the dynamic graph embedding matrix to generate an integrated feature matrix; constructing a multi-objective Lyapunov function based on the integrated feature matrix, generating a robust decision strategy by solving the Hamilton-Jacobi-Bellman equation; and performing multi-dimensional performance monitoring and statistical manifold learning on the robust decision strategy to achieve strategy optimization. The present invention can effectively handle the uncertainty in grid operation and improve the accuracy and reliability of dispatch decisions.

Description

Technical Field

[0001] The present invention belongs to the field of dispatching decision-making, and in particular to a power grid dispatching decision-making method and a large model based on a graph neural network. Background Art

[0002] As a key infrastructure in modern society, the safe and stable operation of the power grid is of great significance to ensuring social and economic development. In recent years, with the large-scale access of renewable energy, increasingly complex load characteristics, and continuous expansion of the power grid structure, the existing power grid dispatching decisions are facing more and more challenges. Especially in the context of deep integration of sources, grids, loads and storage, the operation of the power grid is highly dynamic, strongly coupled and uncertain, and traditional dispatching decision-making methods are difficult to adapt to this complexity. Therefore, using artificial intelligence technology, especially graph neural networks, to improve the intelligence level of power grid dispatching decisions has become a research hotspot in academia and engineering.

[0003] Current research focuses on three aspects: first, grid state estimation based on deep learning, which predicts and evaluates the grid operation state by building a deep neural network model; second, dispatch strategy optimization based on reinforcement learning, which learns the optimal dispatch strategy through the interaction between the intelligent agent and the environment; third, grid topology analysis based on traditional graph theory, which analyzes the grid structure characteristics by extracting graph features. These methods have improved the intelligence level of grid dispatch decisions to a certain extent, but they still have limitations. Traditional deep learning methods are difficult to effectively handle the dynamic topological structure of the power grid; conventional reinforcement learning algorithms have low training efficiency when facing high-dimensional state spaces; and traditional graph theory methods cannot fully utilize the dynamic feature information of the power grid.

[0004] However, the existing technical solutions still face several key technical problems in practical applications: First, at the data processing level, the existing methods are difficult to effectively handle the fusion of multi-source heterogeneous data such as voltage, current, and power, especially when the data distribution changes dynamically, it is difficult to ensure the accuracy of feature extraction; second, in terms of topological representation, the existing graph embedding methods are difficult to simultaneously maintain the local structural characteristics and global topological information of the power grid, resulting in insufficient representation capabilities when facing complex topological changes; third, at the decision optimization level, the existing methods are difficult to effectively handle the impact of system parameter uncertainty, especially in the Markov jump system framework, it is difficult to accurately characterize the dynamic characteristics of state transition; finally, in terms of knowledge transfer, the existing methods lack geometric feature analysis of decision knowledge, and it is difficult to achieve effective generalization of scheduling experience in different scenarios. The existence of these technical problems seriously restricts the level of intelligence of power grid scheduling decisions. Summary of the invention

[0005] The purpose of the invention is to provide a power grid dispatching decision method and a large model based on graph neural network to solve the above-mentioned problems existing in the prior art.

[0006] The technical solution is a grid dispatch decision-making method based on graph neural network, which includes the following steps:

[0007] S1. Collect the original power grid data set, and perform adaptive distribution mapping processing on it to generate a standardized data set; perform multiple recursive decomposition operations on the standardized data set to obtain a multi-scale feature matrix; perform dynamic graph correlation analysis on the multi-scale feature matrix to obtain a dynamic feature matrix and a feature association matrix; based on the dynamic feature matrix and the feature association matrix, perform adaptive multimodal fusion processing to generate a fusion feature tensor; wherein the original power grid data set includes voltage data, current data, power data, frequency data, equipment status data and historical operation data;

[0008] S2. Based on the fused feature tensor and the pre-configured power grid topology matrix, a topology feature matrix and an eigenvalue sequence are generated through spectral decomposition-enhanced topology feature extraction processing; based on the eigenvalue sequence and the pre-stored historical topology change records, change detection is performed on the matrix manifold to obtain a topology change index; based on the topology feature matrix and the fused feature tensor, a multi-relation tensor is constructed and decomposed to obtain a heterogeneous graph structure; based on the heterogeneous graph structure and the topology change index, a spectral-preserving dynamic graph embedding process is performed to output a dynamic graph embedding matrix;

[0009] S3. Based on the dynamic graph embedding matrix and the fused feature tensor, a multi-level feature set is obtained through hierarchical spectrum analysis; Markov jump process analysis is performed on the multi-level feature set to generate a time series feature matrix; the time series feature matrix is ​​input into the sparse graph regularization learning unit to obtain a feature association matrix; based on the time series feature matrix and the feature association matrix, feature fusion processing for optimal transmission enhancement is performed to output an integrated feature matrix;

[0010] S4. Based on the integrated characteristic matrix and the pre-stored operating constraint set, a constraint polynomial set is generated through semi-algebraic set constraint conversion processing; based on the integrated characteristic matrix and the constraint polynomial set, a multi-objective Lyapunov function is constructed; based on the multi-objective Lyapunov function and the constraint polynomial set, a stochastic approximation dynamic programming is performed to obtain a decision strategy and a value function; the decision strategy and the value function are solved by the Hamilton-Jacobi-Bellman equation to generate a robust decision strategy;

[0011] S5. Obtain real-time status data from system feedback, and generate a performance deviation matrix through multi-dimensional performance monitoring based on the real-time status data, robust decision-making strategy and pre-stored historical performance indicators; perform recursive least squares adaptive operations based on the performance deviation matrix and robust decision-making strategy, and output update strategies and parameter estimates; obtain a knowledge base through statistical manifold learning based on parameter estimates and pre-stored historical decision data; perform real-time convex optimization processing based on the update strategy and knowledge base to generate optimization strategies and adjust parameter sets.

[0012] The large-scale model of power grid dispatching decision-making based on graph neural network includes:

[0013] at least one processor; and,

[0014] a memory communicatively connected to at least one of the processors; wherein,

[0015] The memory stores instructions that can be executed by the processor, and the instructions are used to be executed by the processor to implement the power grid dispatch decision method based on graph neural network.

[0016] Beneficial effects: The present invention can accurately capture the dynamic characteristics and topological changes of the power grid, providing a reliable data basis for decision-making; at the same time, it can effectively deal with the uncertainty and disturbance in the operation of the power grid, realize the continuous accumulation of knowledge and real-time optimization of decision-making; improve the accuracy and reliability of power grid dispatching decisions, and also enhance the system's adaptability and anti-interference capabilities, which is of great significance to ensuring the safe and stable operation of the power grid. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 Flow chart of the method of the present invention.

[0018] Figure 2 This is a flow chart of step S1 of the present invention.

[0019] Figure 3 This is a flow chart of step S2 of the present invention.

[0020] Figure 4 This is a flow chart of step S3 of the present invention.

[0021] Figure 5 This is a flow chart of step S4 of the present invention.

[0022] Figure 6 This is a flow chart of step S5 of the present invention. DETAILED DESCRIPTION

[0023] like Figure 1 As shown, this application proposes a power grid dispatch decision method based on graph neural network, including the following steps:

[0024] S1. Collect the original power grid data set, and perform adaptive distribution mapping processing on it to generate a standardized data set; perform multiple recursive decomposition operations on the standardized data set to obtain a multi-scale feature matrix; perform dynamic graph correlation analysis on the multi-scale feature matrix to obtain a dynamic feature matrix and a feature association matrix; based on the dynamic feature matrix and the feature association matrix, perform adaptive multimodal fusion processing to generate a fusion feature tensor; wherein the original power grid data set includes voltage data, current data, power data, frequency data, equipment status data and historical operation data;

[0025] S2. Based on the fused feature tensor and the pre-configured power grid topology matrix, a topology feature matrix and an eigenvalue sequence are generated through spectral decomposition-enhanced topology feature extraction processing; based on the eigenvalue sequence and the pre-stored historical topology change records, change detection is performed on the matrix manifold to obtain a topology change index; based on the topology feature matrix and the fused feature tensor, a multi-relation tensor is constructed and decomposed to obtain a heterogeneous graph structure; based on the heterogeneous graph structure and the topology change index, a spectral-preserving dynamic graph embedding process is performed to output a dynamic graph embedding matrix;

[0026] S3. Based on the dynamic graph embedding matrix and the fused feature tensor, a multi-level feature set is obtained through hierarchical spectrum analysis; Markov jump process analysis is performed on the multi-level feature set to generate a time series feature matrix; the time series feature matrix is ​​input into the sparse graph regularization learning unit to obtain a feature association matrix; based on the time series feature matrix and the feature association matrix, feature fusion processing for optimal transmission enhancement is performed to output an integrated feature matrix;

[0027] S4. Based on the integrated characteristic matrix and the pre-stored operating constraint set, a constraint polynomial set is generated through semi-algebraic set constraint conversion processing; based on the integrated characteristic matrix and the constraint polynomial set, a multi-objective Lyapunov function is constructed; based on the multi-objective Lyapunov function and the constraint polynomial set, a stochastic approximation dynamic programming is performed to obtain a decision strategy and a value function; the decision strategy and the value function are solved by the Hamilton-Jacobi-Bellman equation to generate a robust decision strategy;

[0028] S5. Obtain real-time status data from system feedback, and generate a performance deviation matrix through multi-dimensional performance monitoring based on the real-time status data, robust decision-making strategy and pre-stored historical performance indicators; perform recursive least squares adaptive operations based on the performance deviation matrix and robust decision-making strategy, and output update strategies and parameter estimates; obtain a knowledge base through statistical manifold learning based on parameter estimates and pre-stored historical decision data; perform real-time convex optimization processing based on the update strategy and knowledge base to generate optimization strategies and adjust parameter sets.

[0029] like Figure 2As shown, according to one aspect of the present application, step S1 is further:

[0030] S11. Obtain the voltage data, current data, power data, frequency data of each node in the power grid, as well as the equipment status data of each device and the historical operation data of the system to form an original power grid data set; calculate the multi-level distribution characteristic value based on each data sample in the original power grid data set; perform local density estimation on the multi-level distribution characteristic value to obtain an estimation result; based on the estimation result, dynamically adjust the feature weight coefficient to obtain the weighted feature; use a recursive optimal transmission algorithm to map and transform the weighted feature, and output a standardized data set;

[0031] S12. Based on the standardized data set, a hierarchical recursive decomposition unit is constructed to perform nonlinear transformation and gating operations on the data to obtain transformed data; based on the transformed data, feature components of different frequency levels are calculated; the feature components of each frequency level are combined to generate a multi-scale feature matrix;

[0032] S13. Based on the multi-scale feature matrix, calculate the nonlinear correlation function value between the features; based on the nonlinear correlation function value, construct a feature dynamic association graph through the time-varying structure matrix; perform graph decomposition on the feature dynamic association graph to obtain the main feature components; based on the main feature components, reconstruct the dynamic feature matrix and the feature association matrix;

[0033] S14. Input the dynamic feature matrix and the feature association matrix into the adaptive multimodal fusion processing unit, calculate the feature interaction weight through the adaptive weight generation function; based on the feature interaction weight, fuse the weighted features to obtain the fusion result; perform multimodal integration operation on the fusion result and output the fusion feature tensor.

[0034] In one embodiment of the present application, the original power grid data set D_raw = {real-time measurement data M_t, equipment status data E_t, historical operation data H_t} is obtained, and adaptive distribution mapping (Adaptive Distribution Mapping, ADM) is introduced to construct a multi-level data distribution feature extractor: R(x) = ∑(α_i·φ_i(x)), where φ_i(x) is the feature extraction function of the i-th level; α_i is the adaptive weight coefficient, which is dynamically adjusted through local density estimation; x is the input data sample. Data standardization is achieved through recursive optimal transport (ROT): T(x) = argmin D(R(x), R(y)); where D is the Wasserstein distance; y is the target distribution sample; the output data standardized data set D_norm.

[0035] Based on the standardized dataset D_norm, a hierarchical recursive decomposition unit is constructed through a multi-recursive decomposition network (MRDN): H_k(x) = σ(W_k · x + b_k)ΘG_k(x); where σ is a nonlinear activation function; W_k, b_k are learnable parameters; G_k is the gating function of the k-level; Θ is the Hadamard product; adaptive frequency component extraction is performed: F_i = ∑(H_k(x)·M_i(k)); where M_i is the frequency selection matrix; i represents different frequency levels; and the multi-scale feature matrix T_feat is output.

[0036] Based on the multi-scale feature matrix T_feat, dynamic spectral correlation analysis (DSCA) is performed to construct a feature dynamic association graph: C(i, j) = f(T_feat_i, T_feat_j)·S(i, j); where f is a nonlinear correlation function; S is a time-varying structure matrix. The main feature components are obtained through spectral decomposition: V =Φ(C)·Ψ(C); where Φ is the spectral decomposition operator; Ψ is the feature reconstruction operator; the output data dynamic feature matrix T_dyn and the feature association matrix C.

[0037] In another embodiment of the present application, the process of performing mapping conversion is specifically as follows: feature mapping function F(x) = Σ(αi·K(x, xi) + β·h(x)); where K(x, xi) = exp(-γ||x-xi|| 2 / σ 2 ) is the adaptive kernel function; h(x) = tanh(Wx + b) is the nonlinear transformation function; αi is the sample weight coefficient, which is obtained by local density ρi = Σexp(-||xi-xj|| 2 / dc 2 ) is calculated; β is the balance factor, which is determined by cross-validation; γ is the kernel function parameter, which is adaptively adjusted according to the data distribution; σ is the bandwidth parameter, which is estimated by the median method; W is the weight matrix, which is optimized by back propagation; b is the bias vector, which is updated by gradient descent; x is the input feature vector; xi is the reference sample point; dc is the cutoff distance, which is determined by nearest neighbor statistics.

[0038] Construct a hierarchical recursive decomposition unit: the recursive decomposition function R(X) = H(X) + ΣWi·G(Fi(X)); where H(X) = sigmoid(VX + c) is the nonlinear transformation function; G(Y) = YΘtanh(σ(Y)) is the gating function; Fi(X) =UiX + di is the feature extraction function; Wi is the hierarchical weight matrix, which is adaptively updated through the attention mechanism; V is the transformation matrix, which is determined by gradient optimization; Ui is the extraction matrix, which is designed for different frequency levels; c and di are bias vectors, which are adjusted by back propagation; σ(·) is the activation function; Θ is the Hadamard product; X is the input data matrix.

[0039] Multimodal feature fusion algorithm: Feature fusion function M(F) = Σλi·T(Fi)·S(Fi,Fj); where T(Fi) = softmax(QFiKi T )·Vi is the self-attention transformation; S(Fi,Fj) = exp(-||μi-μj|| 2 / 2θij 2 ) is the modal similarity function; λi is the adaptive weight, which is dynamically adjusted by the information entropy Hi = -Σpi·log(pi); Q, K, V are the query, key, and value matrices; μi is the modal feature mean; θij is the dynamic temperature parameter; Fi is the i-th modal feature; and pi is the feature probability distribution.

[0040] This embodiment achieves high-quality standardization and feature extraction of power grid data through multi-level data processing and feature fusion. Specifically, the original power grid data is first standardized by an adaptive distribution mapping processing method, which can dynamically adjust the feature weights according to the local density of the data, ensuring that data of different dimensions and different distribution characteristics can be processed uniformly. Subsequently, multi-scale features are extracted through multiple recursive decomposition operations, which can capture the changing characteristics of power grid data at different time and space scales. Especially in the dynamic spectrum correlation analysis link, by constructing a feature dynamic association graph and performing spectrum decomposition, not only the time-varying correlation relationship between features is retained, but also the key feature components can be effectively identified. Finally, an adaptive multimodal fusion processing unit is adopted to achieve intelligent fusion of different modal features through adaptive weight generation functions and feature interaction functions. This embodiment improves the accuracy and reliability of subsequent decisions, while reducing the impact of data noise and redundant information on decisions.

[0041] According to one aspect of the present application, step S11 is further:

[0042] S111, receiving voltage data, current data, power data and frequency data of each node in the power grid, calculating the phase angle difference between the nodes; constructing a phase angle correlation matrix based on the phase angle difference; performing data consistency check based on the phase angle correlation matrix, and generating post-check measurement data;

[0043] S112, obtaining device status data of the device, extracting switch status and operating parameters; constructing a device state vector based on the switch status and operating parameters; performing a time series correlation analysis on the device state vector to obtain an analysis result; based on the analysis result, filtering valid status data, and outputting filtered status data;

[0044] S113, reading the historical operation data of the system, extracting the load curve and fault records; calculating the trend component and the periodic component of the time series based on the load curve and the fault records; using an anomaly detection algorithm based on the trend component and the periodic component, identifying and marking the abnormal points in the historical operation data; removing the abnormal points in the historical operation data, and generating preprocessed historical data;

[0045] S114, constructing a multidimensional data tensor based on the verified measurement data, the screened state data and the pre-processed historical data; performing standardization processing on the multidimensional data tensor to obtain a processing result; merging all the processing results to obtain an original power grid data set;

[0046] S115, receiving data samples in the original power grid data set, calculating local density distribution; constructing a hierarchical feature extractor based on the local density distribution; based on the hierarchical feature extractor, adaptively weighting the output of each layer of extractors to obtain multi-level feature values;

[0047] S116. Based on multi-level eigenvalues, a transmission cost matrix is ​​constructed to calculate source distribution and target distribution; based on source distribution and target distribution, the optimal transmission problem is solved recursively to obtain an optimal transmission solution; the optimal transmission solution is iteratively optimized until convergence to generate a standardized data set.

[0048] In one embodiment of the present application, the phase angle correlation analysis algorithm: the phase angle correlation function C(θ) = Σωij·exp(j(θi-θj)) / (N·M); wherein θi is the phase angle of node i; ωij is the node weight coefficient, calculated by ωij = exp(-dij / σ); dij is the electrical distance between nodes; σ is the bandwidth parameter; N is the number of nodes; M is the time window length; j is the imaginary unit.

[0049] Algorithm for time series data preprocessing: sequence decomposition function X(t) = T(t) + S(t) + R(t) + ε(t); where T(t) = Σak·pk(t) is the trend term, pk(t) is the polynomial basis function; S(t) = Σbk·sin(2πkt / τ+φk) is the periodic term; R(t) is the residual term, obtained by wavelet decomposition; ε(t) is the noise term; ak and bk are coefficients; φk is the phase; τ is the basic period.

[0050] This embodiment realizes high-quality collection and preprocessing of raw data of the power grid. By calculating the phase angle difference between nodes and constructing a phase angle correlation matrix, the system can accurately identify and correct outliers in the measurement data. In the equipment status data processing link, the effective status data is screened through time series correlation analysis, which effectively reduces the impact of data noise. For historical operation data, the system realizes effective cleaning and preprocessing of data through time series decomposition and anomaly detection technology. In particular, when constructing the original power grid data set, the standardized processing of multi-dimensional data tensors ensures the unified representation of different types of data. This embodiment improves the accuracy of subsequent feature extraction and decision optimization, and provides a high-quality data basis for power grid dispatching decisions. This is especially important in power grid operation, because accurate raw data is a prerequisite for ensuring safe and stable operation of the power grid.

[0051] According to one aspect of the present application, step S14 is further:

[0052] S141, receiving a dynamic feature matrix, calculating mutual information between feature dimensions; constructing a feature correlation network; determining a local connection pattern based on the network structure; and outputting a feature correlation structure and a connection pattern matrix;

[0053] S142, read the feature association matrix, extract the block diagonal structure of the matrix; perform singular value decomposition on each block; calculate the weight coefficient according to the importance of the singular value; generate the association weight matrix;

[0054] S143, obtaining a feature association structure and a connection mode matrix, constructing an adaptive multimodal fusion processing unit including an adaptive weight generation function; calculating an importance score of a feature combination; dynamically adjusting a weight parameter according to the score; and outputting an adaptive weight tensor;

[0055] S144, import the association weight matrix and adaptive weight tensor, construct a multi-level interaction operator; calculate the high-order association between features; fuse the interaction results of different levels; generate a feature interaction matrix;

[0056] S145, receiving a feature interaction matrix, constructing a modal separator; extracting independent components of each modal feature; calculating a complementarity index between modalities; and outputting a modal feature set;

[0057] S146, obtaining a modal feature set, calculating the information entropy of each modality; determining a fusion weight based on the entropy value; performing a weighted fusion operation; and generating a final fusion feature tensor through iterative optimization.

[0058] In one embodiment of the present application, the mutual information calculation algorithm is: feature mutual information function I(X, Y) = ΣΣp(x, y)·log(p(x, y) / (p(x)p(y))); wherein p(x, y) is a joint probability distribution obtained by kernel density estimation; p(x) and p(y) are marginal probability distributions; x and y are feature vectors; correlation network construction function R(i, j) = I(Xi, Xj) / sqrt(H(Xi)·H(Xj)); H(X) = -Σp(x)·log(p(x)) is information entropy.

[0059] Adaptive weight generation algorithm: weight calculation function W(F) = softmax(η·Φ(F)·Ψ(F)); where Φ(F) = tanh(U·F + b) is the feature transformation function; Ψ(F) = σ(V·F + c) is the importance scoring function; F is the feature matrix; U and V are transformation matrices; b and c are bias vectors; η is the temperature parameter; σ is the sigmoid function.

[0060] Modal separation algorithm: Modal separation function M(X) = argmin||X - ΣDk·Sk||_F + λ·Σ||Sk||* + μ·Ω(D); where Dk is the modal dictionary; Sk is the sparse coding matrix; ||·||_F is the Frobenius norm; ||·||* is the nuclear norm; Ω(D) is the dictionary regularization term; λ and μ are balance parameters; X is the mixed feature matrix.

[0061] This embodiment realizes the adaptive multi-modal fusion of power grid features, which is of great significance to the power grid dispatching decision. First, by calculating the mutual information between feature dimensions and constructing a feature correlation network, the system can accurately capture the potential correlation between the various monitored quantities of the power grid. In the feature correlation matrix processing, the effective identification and extraction of key features are realized through singular value decomposition and importance weight calculation. The adaptive multi-modal fusion processing unit dynamically adjusts the weight parameters of the feature combination, and the system can adaptively adjust the fusion strategy according to the changes in the power grid operation state. In the process of generating the feature interaction matrix, the high-order correlation between features is calculated by multi-level interaction operators, which effectively captures the complex nonlinear relationship in the power grid operation. Finally, through the modal fusion optimization based on information entropy, the system realizes the optimal integration of different modal features. This embodiment improves the accuracy and completeness of the power grid state representation and provides more reliable feature support for dispatching decisions. In the actual operation of the power grid, this feature fusion capability is of great value for accurately identifying potential faults and predicting system fluctuations.

[0062] like Figure 3 As shown, according to one aspect of the present application, step S2 is further:

[0063] S21. Based on the fused feature tensor and the preconfigured power grid topology matrix, calculate the first degree matrix of the topology matrix; based on the power grid topology matrix, construct an adjacency matrix; based on the adjacency matrix and the first degree matrix, calculate the first Laplace operator; use the first Laplace operator and the first degree matrix to standardize the power grid topology matrix to obtain a standardized matrix; perform eigenvalue decomposition on the standardized matrix, extract eigenvalues ​​and eigenvectors, and store them as a topology feature matrix and an eigenvalue sequence, respectively;

[0064] S22, constructing a metric matrix based on the eigenvalue sequence and the pre-stored historical topology change records; calculating the difference statistics of the eigenvalues ​​at adjacent moments based on the metric matrix; generating a Wishart distribution based on the historical topology change records and determining the threshold parameters; analyzing the topology change situation based on the difference statistics and the threshold parameters and generating a topology change index; predicting the topology structure at the next moment based on the topology change index and outputting a predicted topology matrix;

[0065] S23. Based on the topological feature matrix and the fused feature tensor, construct a three-dimensional tensor representing multiple relationships between nodes; perform CP decomposition on the three-dimensional tensor to obtain a factor vector group; based on the factor vector group, reconstruct the relationship representation to obtain a reconstruction result; convert the reconstruction result into a heterogeneous graph structure;

[0066] S24. Construct a dynamic graph embedding processing unit, including a graph structure feature mapping function; input the heterogeneous graph structure and topology change index into the dynamic graph embedding processing unit, and calculate the mapping matrix using the graph Laplacian matrix constraint; based on the topology change index, dynamically update the mapping matrix to obtain an updated mapping matrix; based on the updated mapping matrix, embed the graph structure to generate a dynamic graph embedding matrix.

[0067] In one embodiment of the present application, based on the fused feature tensor F_fused and the current power grid topology matrix A_t, spectral-enhanced topology feature extraction (SETFE) is performed to construct a topology feature matrix: T(A_t) = L·D -1 / 2 ·A_t·D -1 / 2 ; Where L is the Laplace operator; D is the degree matrix. Introduce dynamic spectral decomposition: λ_i(t), v_i(t) = eigv(T(A_t)); where λ_i(t) is the eigenvalue at time t; v_i(t) is the corresponding eigenvector; output data topological feature matrix T_topo and eigenvalue sequence Λ_t.

[0068] Based on the eigenvalue sequence Λ_t and the historical topological change record H_topo, we perform change detection on the matrix manifold (Matrix Manifold Change Detection, MMCD) and construct the topological change detection statistic: S(t) = tr[(Λ_t - Λ_{t-1}) T ·M·(Λ_t - Λ_{t-1})]; where M is the metric matrix; tr represents the trace of the matrix. The threshold is established based on the Wishart distribution: P(S(t) > η) = ∫ η ∞ W_d(n,Σ)dx; where W_d is the Wishart distribution; n is the degree of freedom; Σ is the covariance matrix; output topology change index S_t and predicted topology matrix A_pred.

[0069] Based on the topological feature matrix T_topo and the fused feature tensor F_fused, a tensor decomposition enhanced heterogeneous graph construction (TDEHGC) is performed to construct a multi-relation tensor: R = {r_ijk} ∈ R n×n×k ; where i, j are node indices; k is the relationship type. Relation representation is obtained through CP decomposition: R ≈ ∑_{f=1} F a_fΘb_fΘc_f; where F is the decomposition rank; a_f, b_f, c_f are factor vectors; output heterogeneous graph structure G_het.

[0070] Based on the heterogeneous graph structure G_het and the topological change index S_t, spectral preserving dynamic graph embedding (SPDGE) is performed to construct a spectrum preserving mapping: min ||Φ(G_het)·W - Y||_F + α·tr(W T ·L·W); where Φ is the graph structure feature map; W is the mapping matrix; L is the graph Laplacian matrix; α is the regularization parameter. The dynamic update strategy is: W_t = W_{t-1} + β·▽W·S_t; where β is the learning rate; ▽W is the gradient; the output dynamic graph embedding matrix E_dyn.

[0071] In another embodiment of the present application, a heterogeneous graph structure generation algorithm: a graph structure generation function G(X, E) = P(A)·Φ(X)·Ψ(E); wherein P(A) = softmax(MLP(A)) is an adjacency probability matrix; Φ(X) = ΣviReLU(UiX + bi) is a node feature transformation; Ψ(E) = tanh(WE + c) is an edge feature mapping; A is an adjacency matrix; X is a node feature matrix; E is an edge feature matrix; vi, Ui, W are transformation parameters; bi, c are bias terms; and MLP is a multilayer perceptron.

[0072] Dynamic graph embedding matrix algorithm: Embedding function E(G) = argmin||L - L'||_F +α·tr(Z T LZ) +β·Ω(Z); where L = D - A is the graph Laplacian matrix; L' = D' - f(Z·Z T ) is the reconstruction Laplace matrix; Z is the embedding matrix; D is the degree matrix; A is the adjacency matrix; f(·) is the nonlinear reconstruction function; tr(·) is the trace of the matrix; Ω(Z) is the regularization term; α and β are balance coefficients.

[0073] This embodiment constructs a powerful dynamic graph embedding framework, which realizes efficient characterization and dynamic feature extraction of power grid topology. Through the topological feature extraction processing enhanced by spectral decomposition, the system can accurately capture the essential characteristics of the power grid topology and realize accurate monitoring of topological changes through eigenvalue analysis. In the process of change detection on the matrix manifold, the method of determining threshold parameters by combining Wishart distribution enables the system to accurately identify significant changes in the topological structure. In particular, when constructing heterogeneous graph structures, the CP decomposition method is used to process multi-relation tensors, which not only retains the complex relationship between nodes, but also reduces the computational complexity. Through the spectral-maintained dynamic graph embedding processing unit, the system can achieve rapid response to dynamic changes while maintaining the key features of the topological structure. This embodiment not only improves the accuracy of the representation of the power grid topology, but also improves the system's adaptability to topological changes and prediction accuracy.

[0074] According to one aspect of the present application, step S23 is further:

[0075] S231, based on the topological feature matrix, construct feature grouping based on node type to obtain a node type feature matrix; based on the node type feature matrix, calculate the statistical moment of the intra-group feature; based on the statistical moment, use a polynomial kernel function to calculate the similarity between groups to obtain a type similarity matrix;

[0076] S232, extracting time series pattern features based on the fused feature tensor; calculating the autocorrelation function of the feature sequence based on the time series pattern features; constructing a time dependency relationship based on the autocorrelation function, and generating a time series dependency matrix;

[0077] S233, constructing a multi-relation adjacency tensor based on the node type feature matrix and the type similarity matrix; calculating the eigenvalue distribution of each slice in the multi-relation adjacency tensor; determining the relationship strength based on the eigenvalue distribution, and outputting the relationship strength tensor;

[0078] S234. Based on the temporal dependency matrix and the relationship strength tensor, a three-dimensional relationship representation tensor is constructed; based on the three-dimensional relationship representation tensor, Tucker decomposition is applied to each dimension, a core tensor and a factor matrix are calculated, and a factor vector group is generated;

[0079] S235. Based on the factor vector group, calculate the orthogonality index of the factor vector; based on the orthogonality index, construct a factor importance scoring function to obtain a scoring result; based on the scoring result, screen the main factors in the factor vector group to obtain a core factor matrix;

[0080] S236. Based on the core factor matrix, a heterogeneous relationship reconstruction function is constructed to calculate the multidimensional relationship strength between node pairs; based on the multidimensional relationship strength, different types of relationship representations are extracted; different types of relationship representations are fused to generate fused relationship representations; based on the fused relationship representations, a heterogeneous graph structure is generated through iterative optimization.

[0081] In one embodiment of the present application, a multi-relation tensor construction algorithm: tensor generation function T(A, X) = Σωk·(Ak○Xk); wherein: Ak is the adjacency matrix of the k-th relationship; Xk is the node feature matrix; ωk is the relationship weight, determined by the eigenvalue distribution; ○ is the tensor product; relationship strength calculation function S(i, j, k) = exp(-||fi - fj|| 2 / 2σk 2 ), fi is the node feature vector, and σk is the relationship scale parameter.

[0082] Tucker decomposition algorithm: core tensor calculation function G = X×1U(1)×2U(2)×3U(3); where X is the original tensor; U(n) is the factor matrix; ×n is the n-module tensor product; minimize the objective function L = ||X - G×1U(1)×2U(2)×3U(3)||_F + λ·Ω(U); Ω(U) is the orthogonal constraint term.

[0083] This embodiment constructs a complete heterogeneous graph structure generation framework, which is particularly suitable for processing complex topological relationships in power grids. Through feature grouping based on node type and polynomial kernel function calculation, the system can accurately capture the relationships between different types of power equipment. In the time series pattern feature extraction link, the time dependency is constructed through autocorrelation function analysis, which effectively characterizes the time evolution characteristics of the power grid component state. In particular, in the process of constructing multi-relation adjacency tensors, the system realizes a unified characterization of multiple relationships in the power grid through tensor decomposition and relationship strength analysis. Finally, through Tucker decomposition and factor importance evaluation, the system can extract the most representative network features. This embodiment improves the characterization capability of the power grid topology, can accurately characterize the complex interactions between different types of equipment, and provides more comprehensive topological information support for power grid dispatching decisions. In practical applications, this structured characterization is of great significance for understanding the propagation of power grid faults and optimizing network structures.

[0084] According to one aspect of the present application, step S24 is further:

[0085] S241, receiving a heterogeneous graph structure, extracting the Laplacian matrix of the graph; calculating eigenvectors and eigenvalues; sorting the eigenvectors based on the importance of the eigenvalues; outputting a spectral feature matrix and an eigenvalue sequence;

[0086] S242, read the topology change index, build a time window sequence; calculate the change trend within the window; determine the feature retention threshold based on the trend; generate a dynamic threshold matrix and a trend feature vector;

[0087] S243, obtaining a spectral feature matrix and a eigenvalue sequence, constructing a dynamic graph embedding processing unit including a graph structure feature mapping function; calculating a metric tensor of a mapping space; defining a spectrum-preserving loss function based on the metric tensor; and outputting a feature mapping matrix;

[0088] S244, import the dynamic threshold matrix and trend feature vector, construct the learning rate adaptive function; calculate the gradient update direction; adjust the step size according to the topological change; generate the update parameter set;

[0089] S245, receiving the feature mapping matrix and the update parameter set, constructing a mapping matrix update function; calculating regularization constraint terms; performing gradient projection operations; and outputting an optimized mapping matrix;

[0090] S246. Obtain an optimized mapping matrix and calculate an embedding representation of the graph structure; execute local structure preservation constraints; update the embedding result through iterative optimization; and generate a dynamic graph embedding matrix.

[0091] This embodiment realizes efficient processing of dynamic graph embedding, which is particularly suitable for scenarios with dynamic changes in power grid topology. By extracting the Laplace matrix of the graph and calculating the spectral features, the system can maintain the essential characteristics of the topological structure. In the topology change indicator processing link, accurate monitoring of network structure changes is achieved through dynamic thresholds and trend analysis. The spectrum-preserving mapping mechanism ensures that key topological features are maintained during the dimensionality reduction process through graph structure feature mapping functions and metric tensors. During the mapping matrix update process, the system achieves rapid response to dynamic changes through adaptive learning rates and gradient projections. This embodiment improves the real-time and accuracy of the representation of the power grid topology, can effectively respond to topological changes in power grid operation, and provide timely topological information support for scheduling decisions. In actual power grid operation, this capability is of great value for quickly responding to equipment failures and adjusting network structures.

[0092] like Figure 4 As shown, according to one aspect of the present application, step S3 is further:

[0093] S31. Based on the dynamic graph embedding matrix and the fused feature tensor, a multi-level adjacency matrix and a corresponding second degree matrix are constructed; based on the multi-level adjacency matrix and the corresponding second degree matrix, a second Laplace operator of each level is calculated; based on the second Laplace operator, the features are recursively decomposed using Chebyshev polynomials to obtain decomposition results; the decomposition results of each level are combined to generate a multi-level feature set;

[0094] S32, based on the multi-level feature set, calculating the state transition probability between adjacent moments; based on the state transition probability, constructing a time-varying transition probability matrix; based on the multi-level feature set, constructing a time-varying covariance matrix; based on the transition probability matrix and the time-varying covariance matrix, weighting the features to obtain weighted results; combining the weighted results into a time series feature matrix;

[0095] S33, based on the time series feature matrix, construct an association weight matrix and calculate the graph Laplace matrix; based on the graph Laplace matrix, construct a regularization parameter; based on the regularization parameter, use an alternating direction multiplier method to iteratively optimize the association weight matrix, update the Lagrange multiplier and the auxiliary variable, and generate a feature association matrix;

[0096] S34. Based on the time series feature matrix and the feature correlation matrix, a transmission cost matrix is ​​constructed to calculate the marginal probability of feature distribution; based on the marginal probability, a kernel matrix is ​​generated; based on the kernel matrix, a Sinkhorn iterative algorithm is used to solve the optimal transmission problem to obtain a first solution; based on the first solution, features are fused to output an integrated feature matrix.

[0097] In one embodiment of the present application, based on the dynamic graph embedding matrix E_dyn and the fused feature tensor F_fused, a hierarchical spectral analysis (HSA) is performed to construct a multi-scale Laplacian operator: L_k = I - D_k -1 / 2 ·A_k·D_k -1 / 2 ; where k represents the level; A_k is the adjacency matrix of level k; D_k is the corresponding degree matrix; perform recursive eigendecomposition: F_k =ψ_k(L_k)·F_{k-1}; where ψ_k is the Chebyshev polynomial; F_k is the k-level feature; output the multi-level feature set F_hier.

[0098] Based on the multi-level feature set F_hier, Markov Jump Process Analysis (MJPA) is performed to construct the state transition probability matrix: P(t) = {p_ij(t)} = P(X_{t+1} = j | X_t =i); where X_t represents the state at time t; p_ij(t) is the time-varying transition probability; the time series feature matrix is ​​calculated: ξ(t) = [P(t)·F_hier]·Σ(t); where Σ(t) is the time-varying covariance matrix, and the time series feature matrix ξ_t and the transition probability matrix P_t are output.

[0099] Based on the time series feature matrix ξ_t, sparse graph regularized learning (SGRL) is performed to construct the associated optimization objective: min ||ξ_t - W·ξ_t||_F 2 + λ_1||W||_1 + λ_2tr(ξ_t T ·L·ξ_t); where W is the association weight matrix; L is the graph Laplacian matrix; λ_1, λ_2 are regularization parameters; solve by alternating direction multiplication method, update W: W_{k+1} = argmin L(W, U_k, μ_k); update U: U_{k+1} = S_{λ_1 / ρ}(W_{k+1} + μ_k / ρ); update μ: μ_{k+1} = μ_k + ρ(W_{k+1} - U_{k+1}); output feature association matrix W_corr.

[0100] Based on the time series feature matrix ξ_t and the feature correlation matrix W_corr, the optimal transport problem is constructed through the optimal transport enhanced feature fusion (OTEFF): min_γ∑_{i, j} c_ij·γ_ij st γ1 = μ, γ T 1 = ν; where c_ij is the transmission cost; μ, ν are marginal distributions; solved by the Sinkhorn algorithm: γ_k = diag(u_k)·K·diag(v_k); where K is the kernel matrix; u_k, v_k are scaling vectors; output integrated feature matrix F_int.

[0101] In another embodiment of the present application, a Markov jump process analysis algorithm is used to calculate the state transfer function P(st+1|st) = Σπk·N(μk + Φk·st, Σk); wherein πk is a mixing weight estimated by an EM algorithm; N(·) is a multidimensional Gaussian distribution; μk is a mean vector dynamically updated over time; Φk is a state transfer matrix; Σk is a covariance matrix; st is a state vector at time t; and the jump intensity λij(t) = exp(wij T ·h(t)) / (1 + exp(wij T h(t))) is the jump probability function; h(t) is the hidden state vector; wij is the weight parameter.

[0102] This embodiment implements an efficient feature analysis and fusion processing mechanism. Through hierarchical spectral analysis processing, the system can capture graph structure features from different levels, and realize efficient recursive decomposition of features through Chebyshev polynomials. In the Markov jump process analysis, by constructing time-varying transition probability matrix and covariance matrix, the system can accurately characterize the dynamic evolution law of features. Especially in the sparse graph regularization learning link, the association weight matrix is ​​optimized by alternating direction multiplier method, which not only ensures the sparsity of feature association, but also improves the generalization ability of the model. Finally, the optimal transmission theory is used for feature fusion, and the optimal transmission problem is solved by the Sinkhorn iterative algorithm to achieve the optimal integration of features. This embodiment improves the system's ability to capture the dynamic characteristics of the power grid and provides reliable feature support for subsequent decision-making.

[0103] According to one aspect of the present application, step S32 is further:

[0104] S321. Based on the multi-level feature set, a time series sampling window is constructed; the state distribution within the time series sampling window is calculated to generate a state distribution matrix; based on the state distribution matrix, a kernel density estimation method is used to fit the state probability density to obtain a density function set;

[0105] S322, based on the state distribution matrix, extracting state pairs at adjacent moments and calculating a conditional probability matrix; based on the conditional probability matrix, applying Bayesian estimation to update probability values ​​and generating a jump probability matrix;

[0106] S323, based on the density function set, constructing the jump intensity function between states, calculating the residence time distribution; based on the residence time distribution, fitting the jump time through the exponential distribution, and outputting the jump intensity tensor;

[0107] S324, based on the jump probability matrix and the jump intensity tensor, construct a time-varying transfer probability function, calculate the transfer characteristics of different time scales; based on the transfer characteristics, extract and fuse multi-scale transfer information to generate a transfer probability matrix;

[0108] S325, based on the multi-level feature set, calculating the time series correlation of the feature vector, constructing a covariance estimator; based on the covariance estimator, updating the covariance through a sliding window, and outputting a time-varying covariance matrix;

[0109] S326. Based on the transition probability matrix and the time-varying covariance matrix, a feature weighting function is constructed to calculate the importance weight of the state transition; based on the importance weight, a weighted feature fusion operation is performed to generate a time series feature matrix.

[0110] This embodiment implements the time series feature analysis based on the Markov jump process, which is particularly suitable for the dynamic prediction of the power grid operation status. By constructing a time series sampling window and using the kernel density estimation method, the system can accurately capture the probability distribution characteristics of the power grid state. In the state transition probability calculation link, the accurate modeling of the state transition characteristics is achieved through the Bayesian estimation update and the jump intensity function construction. The processing mechanism of the time-varying transition probability integrates the transfer information of multiple time scales, and the system can accurately characterize the dynamic evolution law of the power grid state. In the feature weighting process, the effective extraction of key time series features is achieved through the time-varying covariance matrix and the state transition importance weight. This embodiment improves the accuracy of power grid state prediction, can effectively capture the sudden events and gradual processes in the power grid operation, and provide reliable prediction support for scheduling decisions. In practical applications, this prediction capability is of great value for preventing system failures, optimizing scheduling strategies, etc.

[0111] like Figure 5 As shown, according to one aspect of the present application, step S4 is further:

[0112] S41, extracting equality constraints and inequality constraints based on the integrated feature matrix and the pre-stored running constraint set; converting the equality constraints and inequality constraints into polynomial forms, and introducing slack variables to construct an optimization problem; solving the optimization problem to obtain a constraint polynomial set and a corresponding slack variable set;

[0113] S42, based on the integrated characteristic matrix and the constraint polynomial set, construct a symmetric positive definite matrix group, calculate the quadratic form and the linear term coefficients; based on the quadratic form and the linear term coefficients, construct and combine a predetermined number of Lyapunov functions, and add a regularization term to obtain a combination result; record the combination result as the objective function, and save the symmetric positive definite matrix group as a Lyapunov matrix set;

[0114] S43, based on the objective function and the constraint polynomial set, calculate the expected value of the state transition and construct an approximate expression of the value function; based on the approximate expression of the value function, update the strategy parameters through the KL divergence constraint, and iteratively optimize to obtain the decision strategy and the corresponding value function;

[0115] S44. Based on the decision strategy and the corresponding value function, construct a cost function containing uncertainty parameters and calculate the gradient of the value function; based on the gradient of the value function, solve the min-max optimization problem; based on the min-max optimization problem, use an iterative method to generate a robust decision strategy and optimized value function that satisfies the HJB equation.

[0116] In one embodiment of the present application, based on the integrated feature matrix F_int and the system operation constraint set C_sys, a semi-algebraic set constraint transformation (SSCT) is performed to construct a polynomial constraint system: h_i(x) ≤ 0, i = 1, ..., m g_j(x) = 0, j = 1, ..., n; where h_i is an inequality constraint polynomial; g_j is an equality constraint polynomial; slack variables are introduced: min_x, s ∑_{i=1} m s_i st h_i(x) + s_i ≤ 0, s_i ≥ 0, g_j(x) = 0; output constraint polynomial set H and slack variable set S.

[0117] Based on the integrated feature matrix F_int and the constraint polynomial set H, a multi-objective Lyapunov function construction (MLFC) is performed to construct the Lyapunov function family: V_k(x) = x T ·P_k·x + q_k T x; where P_k is a symmetric positive definite matrix; q_k is the linear term coefficient; construct a composite objective function: J(x) = ∑_{k=1} Kw_k·V_k(x) + μ·R(x); where w_k is the weight coefficient; R(x) is the regularization term; μ is the regularization parameter; output objective function J and Lyapunov matrix set {P_k},

[0118] Based on the objective function J and the constraint polynomial set H, stochastic approximation dynamic programming (SADP) is performed to construct the value function approximation: V_t(x) = E[J(x_t) +γ·V_{t+1}(x_{t+1})]; where γ is the discount factor; x_t is the state at time t; the strategy update is: π_t(x) =argmin_u {Q_t(x, u) + λ·D_KL(π_t||π_{t-1})}; where Q_t is the state-action value function; D_KL is the KL divergence; λ is the temperature parameter; output the decision strategy π and the value function V.

[0119] Based on the decision strategy π and the value function V, the Hamilton-Jacobi-Bellman equation is solved (HJBSolution Method, HJBSM), and the HJB equation is constructed: Ξ_tV(t, x) = min_u{L(x, u) + ▽V(t, x) T f(x,u)}; where L is the instantaneous cost function; f is the system dynamic equation; perform robustness optimization: π*(x) = argmin_umax_w{L(x,u,w) + ▽V(t,x) T ·f(x,u,w)}; where w is the uncertainty parameter; L(x,u,w) is the cost function with uncertainty; outputs the robust decision strategy π* and the optimized value function V*.

[0120] In another embodiment of the present application, Lyapunov function optimization algorithm: optimization function V(x) = x T ·P(θ)·x + Σαi·Φi(x); where P(θ) = P0 + ΣθiPi is a parameterized positive definite matrix; Φi(x) = log(1 +exp(wi T x + bi)) is a nonlinear basis function; x is the system state vector; θi is the optimization parameter; Pi is the basis matrix; αi is the combination coefficient; wi is the weight vector; bi is the bias term; the constraint condition ▽V(x) T ·f(x) + σ||x|| 2 ≤ 0 ensures stability.

[0121] HJB equation solution algorithm: value function expression V*(x, t) = inf_u sup_w {∫[L(x, u, w) + λ·R(u)]dt + ψ(x(T))}; where L(x, u, w) = x T ·Q·x + u T ·R·u - γ 2 ·w T w is the instantaneous cost; R(u) is the control cost; ψ(x(T)) is the terminal cost; u is the control input; w is the external disturbance; Q and R are weight matrices; γ is the disturbance attenuation rate; λ is the control penalty factor; and T is the terminal time.

[0122] In this embodiment, through semi-algebraic set constraint conversion, the system can convert complex operating constraints into a processable polynomial form. When constructing a multi-objective Lyapunov function, by combining multiple quadratic functions and adding regularization terms, the stability of the system and the diversity of optimization objectives are ensured. In particular, in the stochastic approximation dynamic programming link, the strategy parameters are updated through KL divergence constraints, which not only ensures the exploration ability of the strategy, but also maintains the stability of the decision. When solving the Hamilton-Jacobi-Bellman equation, by constructing a cost function containing uncertainty parameters and iteratively solving the min-max optimization problem, effective processing of system uncertainty is achieved. This embodiment improves the robustness and adaptability of system decision-making, and can effectively deal with various uncertainties in power grid operation.

[0123] According to one aspect of the present application, step S42 is further:

[0124] S421, receiving the integrated feature matrix, constructing a state space representation; calculating the system dynamic features; fitting the local dynamics using a quadratic function; outputting the state dynamic matrix and the local feature set;

[0125] S422, read the constraint polynomial set, extract the constraint boundary conditions; calculate the constraint gradient matrix; construct the feasible domain indicator function based on the gradient; generate the constraint boundary tensor;

[0126] S423, obtaining a state dynamic matrix and a local feature set, constructing a symmetric matrix generating function; calculating an orthogonal basis of eigenvectors; verifying positive definiteness conditions; and outputting a symmetric positive definite matrix group;

[0127] S424, import constraint boundary tensors and symmetric positive definite matrix groups, construct linear term coefficient calculation functions; solve linear programming problems; update coefficients through iterative optimization; generate linear term coefficient matrices;

[0128] S425, receiving a symmetric positive definite matrix group and a linear term coefficient matrix, calculating the weight of each sub-function; constructing an adaptive weight update rule; fusing multiple Lyapunov functions; and outputting a combination function set;

[0129] S426. Obtain a combination function set and construct a regularization function; calculate the gradient penalty term; combine the objective function terms; and generate the final objective function and Lyapunov matrix set.

[0130] This embodiment constructs a complete Lyapunov function optimization framework, which is particularly suitable for the stability control of power grid dispatching. By constructing a state space representation and using a quadratic function to fit local dynamics, the system can accurately characterize the dynamic characteristics of the power grid. In the constraint processing link, the constraint gradient matrix and the feasible domain indicator function are used to achieve accurate modeling of the power grid operation constraints. The generation mechanism of the symmetric positive definite matrix is ​​verified by the eigenvector orthogonal basis and positive definiteness, which ensures the stability of the Lyapunov function. In the process of constructing the combined function, the optimal fusion of multiple Lyapunov functions is achieved through adaptive weight updating and regularization processing. This embodiment improves the stability of power grid dispatching decisions, can effectively ensure the stable operation of the system under various disturbances, and provide theoretical guarantees for dispatching decisions. In actual power grid operation, this stability guarantee is of great significance for maintaining safe system operation and preventing cascading failures.

[0131] According to one aspect of the present application, step S44 is further:

[0132] S441, receiving the decision strategy and value function, constructing the system dynamic equation; calculating the state transfer Jacobian matrix; generating a linearized model based on the Jacobian matrix; outputting the linearized system matrix and the state transfer model;

[0133] S442, read the state transfer model, extract the uncertainty parameter range; construct the probability distribution of parameter changes; calculate the disturbance influence matrix; generate the uncertainty description tensor;

[0134] S443, obtain the linearized system matrix and uncertainty description tensor, construct the instantaneous cost function; calculate the gradient and Hessian matrix of the cost function; solve the optimal control problem; output the cost function matrix;

[0135] S444, import the value function and cost function matrices, construct a value function gradient calculator; apply the finite difference method for discretization; calculate partial derivatives in each direction; generate a value function gradient field;

[0136] S445, receive uncertainty description tensor and value function gradient field, construct min-max optimization problem; design saddle point iteration algorithm; solve game equilibrium strategy; output robust control law;

[0137] S446. Obtain the robust control law and cost function matrix, and construct the HJB equation solver; calculate the optimal trajectory through the characteristic method; iterate the optimization until convergence; generate a robust decision strategy and an optimization value function.

[0138] In one embodiment of the present application, the uncertainty modeling algorithm: parameter distribution function p(w) = Σπi·N(μi,Σi)·I(w∈Wi); wherein πi is a mixing weight; N(μi,Σi) is a Gaussian distribution; I(·) is an indicator function; Wi is a parameter constraint set; a disturbance influence matrix B(w) = J(x,w)·Γ(w); J is a Jacobian matrix; Γ is a disturbance mapping.

[0139] Saddle point iterative algorithm: iterative update function u(k+1) = Proj_U[u(k) - αk·▽uL(u(k),w(k))]; w(k+1) = Proj_W[w(k) + βk·▽wL(u(k),w(k))]; where L is the Lagrangian function; Proj is the projection operator; U and W are constraint sets; αk and βk are step sequences; ▽u and ▽w are gradient operators, u(k+1) is the updated value of the variable u at the k+1th iteration; w(k+1) is the updated value of the variable w at the k+1th iteration.

[0140] This embodiment implements the optimization of robust control strategy based on Hamilton-Jacobi-Bellman (HJB) equation, which is particularly suitable for the control of complex dynamic systems of power grids. By constructing system dynamic equations and linearized models, the system can accurately describe the dynamic characteristics of the power grid. In the uncertainty processing link, by constructing the probability distribution of parameter changes and the disturbance influence matrix, accurate modeling of system uncertainty is achieved. The min-max optimization framework solves the game equilibrium strategy through the saddle point iterative algorithm, and the system can still maintain good control effects in the worst case. In the process of solving the HJB equation, the optimal trajectory is calculated by the characteristic method, and the global optimization of the control strategy is achieved. This embodiment improves the robustness of power grid dispatching decisions, can effectively deal with the uncertainties caused by load fluctuations, renewable energy access, etc., and provides a strong guarantee for the safe and stable operation of the power grid. In practical applications, this robust control capability is of great value for maintaining power grid frequency stability, voltage quality control, etc.

[0141] like Figure 6 As shown, according to one aspect of the present application, step S5 is further:

[0142] S51, obtaining real-time status data fed back by the system, constructing a performance metric tensor based on the real-time status data, the robust decision-making strategy and the pre-stored historical performance indicators, and calculating the weighted norm difference with the reference performance indicator; based on the weighted norm difference, analyzing the spatiotemporal distribution characteristics of the performance deviation, and outputting the performance deviation matrix and the performance trend vector;

[0143] S52, constructing a feature vector and calculating a gain matrix based on the performance deviation matrix and the robust decision strategy; updating the strategy parameters based on the gain matrix to obtain an updated strategy; dynamically adjusting the covariance matrix based on the updated strategy to obtain an adjusted covariance matrix; storing the updated strategy as an updated strategy, and outputting parameter estimates based on the updated strategy and the adjusted covariance matrix;

[0144] S53, based on the parameter estimation value and the pre-stored historical decision data, construct the conditional probability distribution in the parameter space, and calculate the Fisher information metric; based on the Fisher information metric, construct and solve the geodesic equation to obtain the learning result; save the learning result as a knowledge base, and generate the corresponding generalization mapping;

[0145] S54. Based on the update strategy and knowledge base, a local convex approximation problem is constructed to generate a constraint matrix; based on the constraint matrix, the local convex approximation problem is iteratively solved using a fast gradient method to obtain a second solution; a feasibility projection is performed on the second solution, and finally an optimization strategy and an adjustment parameter set are output.

[0146] In one embodiment of the present application, based on the robust decision strategy π*, the real-time system state x_t and the historical performance indicator H_perf, a performance metric tensor is constructed through multi-dimensional performance monitoring (MPM): Ψ(t) = {ψ_ijk(t)} ∈ R p×q×r ; where i represents the performance indicator dimension; j represents the time scale; k represents the spatial region; ψ_ijk is the corresponding performance metric. Perform performance deviation analysis: Δ(t) = ||Ψ(t) - Ψ_ref||_W; where Ψ_ref is the reference performance indicator; ||·||_W is the weighted norm; output the performance deviation matrix Δ_perf and the performance trend vector τ_t.

[0147] Based on the performance deviation matrix Δ_perf and the robust decision strategy π*, the parameter update equation is constructed through recursive least squares adaptation (RLSA): θ_t = θ_{t-1} +K_t(y_t - φ_t T·θ_{t-1}); where θ_t is the policy parameter; K_t is the gain matrix; φ_t is the eigenvector. Update the gain matrix: K_t = P_{t-1}·φ_t(λ + φ_t T ·P_{t-1}·φ_t) -1 ;P_t = (I - K_t·φ_t T )·P_{t-1} / λ; where λ is the forgetting factor; P_t is the covariance matrix; output update strategy π_new and parameter estimate θ_t.

[0148] Based on the parameter estimate θ_t and the historical decision data D_hist, the parameter manifold is constructed through statistical manifold learning (SML): M = {p(x|θ) : θ ∈ Π}; where p(x|θ) is the conditional probability distribution; Π is the parameter space. Calculate the geodesic distance: d(θ_1, θ_2) = inf_γ∫0 1 sqrt(g_ij(γ(t))·dγ i / dt·dγ j / dt)dt; where g_ij is the Fisher information measure; γ(t) is the path in the parameter space; the output data knowledge base K_base and the generalization map Φ_gen.

[0149] Based on the update strategy π_new and the knowledge base K_base, a local convex approximation is constructed through real-time convex optimization (RCO): min_u f_0(u) + ∑ i=1 m f_i(u) st Au = b; Gu≤h; where f_i is a convex function; A, G are constraint matrices. Use the fast gradient method to solve: u_{k+1} = P_Ω(u_k - α_k▽f(u_k)); where P_Ω is the projection operator; α_k is the step size; Ω is the feasible set; output the optimization strategy π_opt and the adjustment parameter set_adj.

[0150] In another embodiment of the present application, a statistical manifold learning algorithm is used to construct and solve the geodesic equation: 2 x μ / dt 2 + Γμνρ(dx ν / dt)(dx ρ / dt) = 0; where Γμνρ is the Kirschner symbol, through Γμνρ = (1 / 2)g μσ(Ξ*νgσρ +Ξ*ρgσν -Ξ*σgνρ) calculation; gμν is Fisher information metric tensor; g μ σ is the inverse of the metric tensor; x μ is the manifold coordinate; t is the parameter; Ξ*ν represents the partial derivative with respect to the νth coordinate; knowledge representation K(x, y) = exp(-d_M(x, y) 2 / 2σ 2 ), where d_M(x, y) is the geodesic distance.

[0151] This embodiment constructs an adaptive performance monitoring and optimization framework, which is particularly suitable for real-time optimization of power grid dispatching decisions. Through multi-dimensional performance monitoring processing, the system can capture the deviation between the power grid operation status and the expected performance in real time, and accurately predict the performance trend by constructing a time-varying covariance matrix. In the recursive least squares adaptive operation link, the system realizes real-time adjustment of the decision-making strategy by dynamically updating the gain matrix and the covariance matrix. By introducing the statistical manifold learning unit, by constructing the Fisher information metric in the parameter space and solving the geodesic equation, the system can realize effective generalization of knowledge while keeping important features unchanged. Finally, through real-time convex optimization processing, the system can quickly solve local optimization problems and perform feasibility projection, ensuring the real-time and feasibility of the optimization results. This embodiment improves the real-time performance of power grid dispatching decisions and ensures the stability and reliability of the system in the face of complex operating environments.

[0152] According to one aspect of the present application, step S51 is further:

[0153] S511, receiving robust decision-making strategies and real-time state data, constructing a multidimensional state observer; calculating real-time values ​​of key performance indicators; extracting time series features; and outputting a performance indicator matrix;

[0154] S512, read historical performance indicators, extract data fragments of different time scales; calculate statistical characteristic parameters; build a performance benchmark model; generate a reference performance tensor;

[0155] S513, obtaining a performance indicator matrix, constructing a region segmentation function; calculating a spatial correlation measure; dividing the monitoring area based on the correlation; outputting a spatial partition map and a regional feature matrix;

[0156] S514, importing the reference performance tensor and the regional feature matrix, constructing a performance measurement calculation function; calculating the normalized weight of each dimension; generating a weighted norm calculation rule; and outputting a weight coefficient set;

[0157] S515, receiving a performance indicator matrix and a set of weight coefficients, calculating weighted deviations from reference values; constructing a spatiotemporal deviation analysis model; evaluating the statistical significance of the deviations; and generating a performance deviation matrix;

[0158] S516. Obtain the performance deviation matrix and spatial partition mapping, construct a trend analysis function; calculate local and global change trends; integrate multi-scale trend information; and output a performance trend vector.

[0159] In this embodiment, the system can comprehensively monitor the operation status of the power grid through multi-dimensional state observers and real-time calculation of key performance indicators. In the performance benchmark modeling stage, dynamic updating of performance standards is achieved through multi-time scale data analysis and statistical feature extraction. The spatial correlation analysis mechanism enables the system to identify the spatial distribution characteristics of performance problems through regional segmentation and correlation measurement. In the performance deviation analysis process, accurate positioning of performance anomalies is achieved through weighted norm calculation and spatiotemporal deviation analysis. This embodiment improves the reliability of power grid dispatching decisions, can timely discover and locate system performance problems, and provide effective feedback for real-time optimization of dispatching strategies. In actual power grid operation, this performance monitoring capability is of great significance for ensuring power supply quality and improving system efficiency.

[0160] According to one aspect of the present application, step S53 is further:

[0161] S531, receiving parameter estimates, constructing a probability density function of the parameter space; calculating a kernel function representation of the conditional probability distribution; generating a high-dimensional feature space through kernel function transformation; and outputting a conditional density matrix;

[0162] S532, read historical decision data, extract decision sample sequence; calculate local distance between samples; construct nearest neighbor relationship graph; generate sample association network and distance matrix;

[0163] S533, obtain the conditional density matrix and distance matrix, construct the Fisher information metric calculator; calculate the natural gradient between probability distributions; evaluate the information geometry characteristics; output the Fisher information matrix;

[0164] S534, import the sample association network and Fisher information matrix, construct the geodesic equation; calculate the shortest path on the manifold; solve the variational problem to obtain the optimal path; generate the geodesic path set;

[0165] S535, receiving a set of geodetic paths, constructing a knowledge extraction function; calculating geometric invariants of the paths; extracting key decision patterns; and outputting a knowledge representation matrix;

[0166] S536. Obtain the knowledge representation matrix and the Fisher information matrix, construct a generalized mapping function; calculate the regularization constraints of the mapping; optimize the mapping parameters until convergence; generate the final knowledge base and generalized mapping.

[0167] This embodiment realizes knowledge learning and generalization based on statistical manifolds, which is particularly suitable for the accumulation and reuse of power grid dispatching experience. By constructing the conditional probability density and kernel function representation of the parameter space, the system can effectively extract the essential characteristics of decision-making knowledge. In the construction link of Fisher information metric, the geometric representation of decision-making knowledge is realized by calculating the natural gradient between probability distributions. The solution mechanism of geodesic equations calculates the shortest path on the manifold, and the system can find the optimal knowledge transfer path. In the process of knowledge extraction and generalization, efficient generalization of decision-making knowledge is achieved through geometric invariant extraction and regularization constraints. This embodiment improves the adaptability and generalization ability of power grid dispatching decisions, can effectively transfer historical successful experiences to new scenarios, and provide knowledge support for dispatching decisions. In practical applications, this knowledge learning ability is of great value in improving dispatching efficiency and reducing decision-making errors.

[0168] In another embodiment of the present application, the graph neural network processing module in step S2 also includes a graph attention mechanism, a temporal graph convolutional network and a multi-head attention mechanism. The graph attention mechanism weights and aggregates neighbor node information by calculating the attention coefficient between nodes. In the specific implementation, the attention score is calculated for each pair of connected nodes, then normalized by the softmax function, and finally feature aggregation is performed based on the attention weight. The temporal graph convolutional network processing realizes the modeling of dynamic topological structures by combining the spatial structure and temporal evolution characteristics of the graph. The network first extracts temporal features, then analyzes the evolution law of the graph structure over time, and finally integrates information from multiple time steps to obtain the final output. The multi-head attention mechanism learns node features from different representation subspaces by running multiple attention layers in parallel, and finally averages the outputs of multiple attention heads to obtain the final node representation.

[0169] In another embodiment of the present application, step S4 also includes: a trajectory condition generation mechanism, a causal attention module and a return-condition-action prediction framework. The trajectory condition generation mechanism processes the historical state sequence through a converter encoder and generates condition information based on the target return using a multi-layer perceptron predictor. The causal attention module implements a multi-head attention mechanism and ensures that the current decision depends only on historical information through a causal mask. The return-condition-action prediction framework first encodes the state, then processes the state, action and return sequence through a causal converter, and finally predicts the next action. By combining the spatial representation capability of the graph neural network and the sequence modeling capability of the decision converter, the performance of the power grid dispatching decision is improved. In terms of graph structure representation, the multi-head attention mechanism can capture different types of node relationships, and the temporal graph convolution effectively models dynamic changes. In terms of decision generation, causal attention ensures the temporal consistency of the decision, and the trajectory condition generation mechanism provides precise control of the decision process. The synergy of these mechanisms enables the system to better adapt to complex scenarios and dynamic changes in power grid operation.

[0170] According to one aspect of the present application, a grid dispatch decision method based on a graph neural network includes a training process and a decision process:

[0171] In the training process, a large number of trajectories (episodes) are generated using expert strategies or random strategies, and a data set consisting of each complete trajectory is collected. Each trajectory contains the reward, state, and action sequence (return1, s1, a1, return2, s2, a2, ..., returnn) in the episode. The input data is preprocessed by the data processing module for state and action, and the data is converted into a format acceptable to the neural network. The state features are extracted through the graph neural network state encoding module to generate a state encoding sequence. The action encoding module processes the action information and generates action encoding. In the decision transformer module, the state, action, and reward sequence are input into the model to predict the next optimal action. The autoregressive "next step prediction" loss function is used, which mainly supervises the action output. Specifically, for the three continuous tokens composed of the action, the discrete part uses the cross entropy loss and the continuous part uses the MSE loss. During training, the error calculated by the loss function is back-propagated to the model parameters, and the model parameters are continuously adjusted through the optimizer (such as Adam) to minimize the loss function.

[0172] In the decision-making process, the environment is initialized: the environment feeds back the initial state s1. The initial defined reward and state are passed through the data processing module and the encoding module to generate an input sequence, which is handed over to the Transformer to generate the first part of the action logic value (logits). The complete action triplet is iteratively generated by sampling (or greedy), and the triplet is handed over to the action decoding module to generate the environment definition action. The environment receives the action, updates the state at the next moment, and returns the reward and the state at the next moment. The reward of the previous step is subtracted from the reward to obtain the reward at this moment, and the encoded sequence is obtained by similar processing with the state. The subsequent decision uses the key-value cache (kv cache) to reduce the amount of calculation during decision-making. At the same time, the action part is generated. Iterate the interaction until the environment reaches the termination state.

[0173] According to one aspect of the present application, a large model for power grid dispatching decision-making based on a graph neural network includes a data preprocessing module, a graph neural network state encoding module, an action encoding module, a reward encoding module, a Decision Transformer module, and a dispatching output module. The data processing module is used to receive the original state and action input, and convert the input into a format that conforms to the neural network structure. The specific implementation includes: converting each independent feature represented by the original state (usually a key-value pair representation) into the format required by the graph neural network, that is, aggregating node features and constructing an adjacency matrix between nodes. Due to the complex structure of the action space, the action is modeled as a triple (a, o, Δp). Among them: a represents the type of action, one of ; the content of o is the number of the object acted on by the discrete action, and the continuous action is an additional number; the Δp part is the adjustment change of the active power of each generator set, and the discrete action is all zero.

[0174] The graph neural network state encoding module uses a heterogeneous graph neural network (HGNN) to learn the original features of each node and obtain a feature representation of a unified dimension. The specific operation is: calculate the corresponding abstract features for each node, and the output shape is n×d, where n is the number of nodes and d is the feature dimension. Aggregate the features of nodes of the same type to generate a coding sequence as the state coding sequence.

[0175] The action encoding module encodes the three parts of the action separately, including: discrete part encoding: discrete actions are encoded using the embedding layer; continuous part encoding: continuous variables are encoded through linear mapping; action decoding: hidden states are mapped to corresponding dimensions, the discrete part outputs logits, and the continuous part outputs action values ​​(which need to be trimmed to stay within a reasonable range). At the same time, in order to ensure that the generated actions are meaningful, unreasonable logits need to be marked and the continuous part of discrete actions needs to be zeroed.

[0176] The return encoding module linearly maps the return scalar value to the same dimension as other sequences.

[0177] The Decision Transformer module is mainly based on the GPT-2 structure. It inputs and processes the reward, state, and action sequence, adopts the autoregressive "next step prediction" loss function, and only performs supervised learning on the action output during training to improve the system's prediction accuracy for the next scheduling action.

[0178] The scheduling output module converts the action triples generated by the action encoding module back into the action form defined by the environment and interacts with the environment so that the generated scheduling strategy can dynamically respond to environmental changes.

[0179] In terms of model architecture, this embodiment introduces the architecture of Decision Transformer into the field of power grid dispatching, which can be easily applied to power grid dispatching scenarios in different regions or different scales, and quickly fine-tune the model according to specific needs and constraints to adapt to the new environment. In terms of data preprocessing, the graph neural network is used to extract the features of the power grid state in the Decision Transformer, so that the model subject can obtain the power grid topology information. At the same time, due to the characteristics of the graph neural network, the model can be trained with data from power systems of different scales, so that the model can be generalized and make decisions in different environments. For the processing of actions, the action of the power grid is modeled as a ternary sequence, so that the sequence modeling capability of the Transformer can be used to model and generate actions in complex action spaces. Through the system of this embodiment, power grid dispatching can achieve efficient and flexible adaptive control in complex scenarios, thereby improving the stability and efficiency of the power grid.

[0180] According to another aspect of the present application, a large model for power grid dispatching decision-making based on graph neural network includes:

[0181] at least one processor; and,

[0182] a memory communicatively connected to at least one of the processors; wherein,

[0183] The memory stores instructions that can be executed by the processor, and the instructions are used to be executed by the processor to implement the power grid scheduling decision method based on graph neural network described in any of the above embodiments.

[0184] The present invention constructs a complete power grid dispatching decision system based on graph neural network, realizing the full process intelligence from data processing, feature extraction, decision optimization to performance monitoring. First, through multimodal data fusion and dynamic graph embedding technology, the system can accurately capture the dynamic characteristics and topological changes of the power grid, and provide a reliable data basis for decision-making. Secondly, based on Markov jump process and Lyapunov optimization theory, the system constructs a robust decision-making framework that can effectively deal with uncertainties and disturbances in power grid operation. Thirdly, by introducing statistical manifold learning and real-time convex optimization, the system realizes the continuous accumulation of knowledge and real-time optimization of decision-making. The present invention not only improves the accuracy and reliability of power grid dispatching decisions, but also enhances the system's adaptability and anti-interference ability, which is of great significance to ensuring the safe and stable operation of the power grid.

[0185] The preferred embodiments of the present invention are described in detail above; however, the present invention is not limited to the specific details in the above embodiments. Within the technical concept of the present invention, various equivalent transformations can be made to the technical solutions of the present invention, and these equivalent transformations all belong to the protection scope of the present invention.

Claims

1. A power grid dispatching decision-making method based on graph neural network, characterized in that: The steps include: S1. Collecting the original power grid data set, and performing adaptive distribution mapping processing on it to generate a standardized data set; performing multiple recursive decomposition operations on the standardized data set to obtain a multi-scale feature matrix; Perform dynamic graph correlation analysis on the multi-scale feature matrix to obtain the dynamic feature matrix and feature correlation matrix; Based on the dynamic feature matrix and feature correlation matrix, adaptive multi-modal fusion processing is performed to generate a fused feature tensor; The original power grid data set includes voltage data, current data, power data, frequency data, equipment status data and historical operation data; S2. Based on the fused feature tensor and the pre-configured power grid topology matrix, a topology feature matrix and an eigenvalue sequence are generated through spectral decomposition-enhanced topology feature extraction processing; based on the eigenvalue sequence and the pre-stored historical topology change records, change detection is performed on the matrix manifold to obtain a topology change index; Based on the topological feature matrix and fused feature tensor, a multi-relation tensor is constructed and decomposed to obtain a heterogeneous graph structure; Based on the heterogeneous graph structure and topological change indicators, a spectrum-preserving dynamic graph embedding process is performed to output a dynamic graph embedding matrix. S3, based on the dynamic graph embedding matrix and the fused feature tensor, a multi-level feature set is obtained through hierarchical spectrum analysis; Perform Markov jump process analysis on multi-level feature sets to generate a time series feature matrix; Input the time series feature matrix into the sparse graph regularization learning unit to obtain the feature association matrix; Based on the time series feature matrix and feature correlation matrix, perform feature fusion processing for optimal transmission enhancement and output the integrated feature matrix; S4. Based on the integrated characteristic matrix and the pre-stored operating constraint set, a constraint polynomial set is generated through semi-algebraic set constraint conversion processing; based on the integrated characteristic matrix and the constraint polynomial set, a multi-objective Lyapunov function is constructed; based on the multi-objective Lyapunov function and the constraint polynomial set, a stochastic approximation dynamic programming is performed to obtain a decision strategy and a value function; the decision strategy and the value function are solved by the Hamilton-Jacobi-Bellman equation to generate a robust decision strategy; S5. Obtain real-time status data fed back by the system, and generate a performance deviation matrix through multi-dimensional performance monitoring processing based on the real-time status data, robust decision-making strategy and pre-stored historical performance indicators; Based on the performance deviation matrix and the robust decision strategy, a recursive least squares adaptive operation is performed to output an update strategy and parameter estimates; Based on parameter estimates and pre-stored historical decision data, a knowledge base is obtained through statistical manifold learning; Based on the update strategy and knowledge base, real-time convex optimization processing is performed to generate optimization strategies and adjust parameter sets.

2. The power grid dispatching decision method based on graph neural network according to claim 1 is characterized in that: Step S1 is further as follows: S11, obtaining voltage data, current data, power data, frequency data of each node in the power grid, as well as device status data of each device and historical operation data of the system to form an original power grid data set; based on each data sample in the original power grid data set, calculating a multi-level distribution characteristic value; Perform local density estimation on multi-level distribution feature values ​​to obtain estimation results; based on the estimation results, dynamically adjust the feature weight coefficient to obtain weighted features; The recursive optimal transfer algorithm is used to map and transform the weighted features and output a standardized data set; S12. Based on the standardized data set, a hierarchical recursive decomposition unit is constructed to perform nonlinear transformation and gating operations on the data to obtain transformed data; Based on the transformed data, characteristic components at different frequency levels are calculated; Combine the feature components of each frequency level to generate a multi-scale feature matrix; S13. Based on the multi-scale feature matrix, calculate the nonlinear correlation function value between the features; based on the nonlinear correlation function value, construct a feature dynamic association graph through the time-varying structure matrix; perform graph decomposition on the feature dynamic association graph to obtain the main feature components; based on the main feature components, reconstruct the dynamic feature matrix and the feature association matrix; S14, inputting the dynamic feature matrix and the feature association matrix into the adaptive multimodal fusion processing unit, and calculating the feature interaction weight through the adaptive weight generation function; Based on the feature interaction weights, the weighted features are fused to obtain the fusion result; Perform multimodal integration operations on the fusion results and output the fused feature tensor.

3. The power grid dispatching decision method based on graph neural network according to claim 2 is characterized in that: Step S2 is further as follows: S21, calculating the first degree matrix of the topology matrix based on the fused feature tensor and the preconfigured power grid topology matrix; Based on the power grid topology matrix, an adjacency matrix is ​​constructed; Calculate the first Laplacian operator based on the adjacency matrix and the first degree matrix; The power grid topology matrix is ​​normalized by using the first Laplace operator and the first degree matrix to obtain a normalized matrix; Perform eigenvalue decomposition on the standardized matrix, extract eigenvalues ​​and eigenvectors, and store them as topological characteristic matrix and eigenvalue sequence respectively; S22, constructing a metric matrix based on the eigenvalue sequence and the pre-stored historical topology change records; Based on the metric matrix, the difference statistics of the eigenvalues ​​at adjacent moments are calculated; based on the historical topological change records, the Wishart distribution is generated and the threshold parameters are determined; Based on the difference statistics and threshold parameters, the topology change is analyzed and the topology change index is generated; based on the topology change index, the topology structure at the next moment is predicted and the predicted topology matrix is ​​output; S23. Based on the topological feature matrix and the fused feature tensor, a three-dimensional tensor representing various relationships between nodes is constructed; the three-dimensional tensor is subjected to CP decomposition to obtain a factor vector group; Based on the factor vector group, the relationship representation is reconstructed to obtain the reconstruction result; the reconstruction result is converted into a heterogeneous graph structure; S24, constructing a dynamic graph embedding processing unit, including a graph structure feature mapping function; The heterogeneous graph structure and topology change indicators are input into the dynamic graph embedding processing unit, and the mapping matrix is ​​calculated using the graph Laplacian matrix constraints; Based on the topology change index, the mapping matrix is ​​dynamically updated to obtain an updated mapping matrix; Based on the updated mapping matrix, the graph structure is embedded to generate a dynamic graph embedding matrix.

4. The power grid dispatching decision method based on graph neural network according to claim 3 is characterized in that: Step S3 is further as follows: S31. Based on the dynamic graph embedding matrix and the fused feature tensor, construct a multi-level adjacency matrix and a corresponding second degree matrix; based on the multi-level adjacency matrix and the corresponding second degree matrix, calculate the second Laplacian operator of each level; Based on the second Laplace operator, the Chebyshev polynomial is used to recursively decompose the features to obtain the decomposition results; the decomposition results of each level are combined to generate a multi-level feature set; S32, calculating the state transition probability between adjacent time points based on the multi-level feature set; Based on the state transition probability, a time-varying transition probability matrix is ​​constructed; based on the multi-level feature set, a time-varying covariance matrix is ​​constructed; Based on the transition probability matrix and the time-varying covariance matrix, the features are weighted to obtain weighted results; the weighted results are combined into a time series feature matrix; S33. Based on the time series feature matrix, construct an association weight matrix and calculate the graph Laplacian matrix; Based on the graph Laplacian matrix, construct the regularization parameters; Based on the regularization parameter, the alternating direction multiplier method is used to iteratively optimize the association weight matrix, update the Lagrange multipliers and auxiliary variables, and generate the feature association matrix; S34, based on the time series feature matrix and the feature association matrix, construct a transmission cost matrix, calculate the marginal probability of feature distribution; based on the marginal probability, generate a kernel matrix; Based on the kernel matrix, the Sinkhorn iterative algorithm is used to solve the optimal transmission problem and obtain a first solution result; based on the first solution result, the features are fused and an integrated feature matrix is ​​output.

5. The power grid dispatching decision method based on graph neural network according to claim 4 is characterized in that: Step S4 is further as follows: S41, extracting equality constraints and inequality constraints based on the integrated feature matrix and the pre-stored running constraint set; converting the equality constraints and inequality constraints into polynomial forms, and introducing slack variables to construct an optimization problem; solving the optimization problem to obtain a constraint polynomial set and a corresponding slack variable set; S42, based on the integrated characteristic matrix and the set of constraint polynomials, construct a symmetric positive definite matrix group, and calculate the quadratic form and linear term coefficients; Based on the quadratic form and the linear term coefficients, a predetermined number of Lyapunov functions are constructed and combined, and a regularization term is added to obtain a combination result; the combination result is recorded as the objective function, and the symmetric positive definite matrix group is saved as a Lyapunov matrix set; S43, based on the objective function and the constraint polynomial set, calculating the expected value of the state transition, and constructing an approximate expression of the value function; Based on the approximate expression of the value function, the policy parameters are updated through the KL divergence constraint, and the decision strategy and the corresponding value function are obtained through iterative optimization; S44. Based on the decision strategy and the corresponding value function, construct a cost function containing uncertainty parameters and calculate the gradient of the value function; based on the gradient of the value function, solve the min-max optimization problem; based on the min-max optimization problem, use an iterative method to generate a robust decision strategy and optimized value function that satisfies the HJB equation.

6. The power grid dispatching decision method based on graph neural network according to claim 5 is characterized in that: Step S5 is further as follows: S51, obtaining real-time status data fed back by the system, constructing a performance metric tensor based on the real-time status data, a robust decision-making strategy, and pre-stored historical performance indicators, and calculating a weighted norm difference with a reference performance indicator; Based on the weighted norm difference, the spatiotemporal distribution characteristics of performance deviation are analyzed, and the performance deviation matrix and performance trend vector are output; S52, constructing a feature vector and calculating a gain matrix based on the performance deviation matrix and the robust decision strategy; Based on the gain matrix, update the strategy parameters to obtain the updated strategy; Based on the updated strategy, dynamically adjust the covariance matrix to obtain an adjusted covariance matrix; store the updated strategy as an updated strategy, and output parameter estimates based on the updated strategy and the adjusted covariance matrix; S53, based on the parameter estimation value and the pre-stored historical decision data, construct the conditional probability distribution in the parameter space, and calculate the Fisher information metric; based on the Fisher information metric, construct and solve the geodesic equation to obtain the learning result; Save the learning results as a knowledge base and generate the corresponding generalization mapping; S54. Based on the update strategy and knowledge base, a local convex approximation problem is constructed to generate a constraint matrix; Based on the constraint matrix, the fast gradient method is used to iteratively solve the local convex approximation problem to obtain the second solution result; the feasibility projection is performed on the second solution result, and finally the optimization strategy and adjustment parameter set are output.

7. The power grid dispatching decision method based on graph neural network according to claim 6 is characterized in that: Step S11 is further as follows: S111, receiving voltage data, current data, power data and frequency data of each node in the power grid, calculating the phase angle difference between the nodes; and constructing a phase angle correlation matrix based on the phase angle difference; Based on the phase angle correlation matrix, data consistency verification is performed to generate verified measurement data; S112, obtaining device status data of the device, extracting switch status and operating parameters; constructing a device state vector based on the switch status and operating parameters; performing a time series correlation analysis on the device state vector to obtain an analysis result; Based on the analysis results, filter the valid status data and output the filtered status data; S113, reading the historical operation data of the system, extracting the load curve and the fault record; and calculating the trend component and the period component of the time series based on the load curve and the fault record; Based on the trend component and the periodic component, anomaly detection algorithms are used to identify and mark abnormal points in historical operation data; Eliminate abnormal points in historical operation data to generate pre-processed historical data; S114, constructing a multidimensional data tensor based on the verified measurement data, the screened state data and the pre-processed historical data; Perform standardization processing on the multi-dimensional data tensor to obtain processing results; merge all processing results to obtain the original power grid data set; S115, receiving data samples in the original power grid data set, calculating local density distribution; and constructing a hierarchical feature extractor based on the local density distribution; Based on the hierarchical feature extractor, the output of each layer of extractor is adaptively weighted to obtain multi-level feature values; S116. Based on multi-level eigenvalues, a transmission cost matrix is ​​constructed to calculate source distribution and target distribution; based on source distribution and target distribution, the optimal transmission problem is solved recursively to obtain an optimal transmission solution; the optimal transmission solution is iteratively optimized until convergence to generate a standardized data set.

8. The grid dispatch decision method based on graph neural network according to claim 6 is characterized in that: Step S23 is further as follows: S231, based on the topological feature matrix, construct feature grouping based on node type to obtain a node type feature matrix; based on the node type feature matrix, calculate the statistical moment of the intra-group feature; based on the statistical moment, use a polynomial kernel function to calculate the similarity between groups to obtain a type similarity matrix; S232, extracting temporal pattern features based on the fused feature tensor; Based on the time series pattern characteristics, the autocorrelation function of the characteristic sequence is calculated; Based on the autocorrelation function, the time dependency relationship is constructed and the time series dependency matrix is ​​generated; S233, constructing a multi-relation adjacency tensor based on the node type feature matrix and the type similarity matrix; Compute the eigenvalue distribution of each slice in the multi-relation adjacency tensor; Based on the eigenvalue distribution, determine the relationship strength and output the relationship strength tensor; S234. Based on the temporal dependency matrix and the relationship strength tensor, a three-dimensional relationship representation tensor is constructed; based on the three-dimensional relationship representation tensor, Tucker decomposition is applied to each dimension, a core tensor and a factor matrix are calculated, and a factor vector group is generated; S235. Based on the factor vector group, calculate the orthogonality index of the factor vector; based on the orthogonality index, construct a factor importance scoring function to obtain a scoring result; Based on the scoring results, the main factors in the factor vector group are screened to obtain the core factor matrix; S236. Based on the core factor matrix, a heterogeneous relationship reconstruction function is constructed to calculate the multidimensional relationship strength between node pairs; Extract different types of relationship representations based on multi-dimensional relationship strength; Fusing different types of relation representations to generate fused relation representations; Based on the fused relationship representation, a heterogeneous graph structure is generated through iterative optimization.

9. The power grid dispatching decision method based on graph neural network according to claim 6 is characterized in that: Step S32 is further as follows: S321. Based on the multi-level feature set, a time series sampling window is constructed; the state distribution within the time series sampling window is calculated to generate a state distribution matrix; based on the state distribution matrix, a kernel density estimation method is used to fit the state probability density to obtain a density function set; S322, based on the state distribution matrix, extracting state pairs at adjacent moments and calculating a conditional probability matrix; Based on the conditional probability matrix, Bayesian estimation is applied to update the probability values ​​and generate a jump probability matrix; S323, constructing a jump intensity function between states based on the density function set, and calculating the residence time distribution; Based on the residence time distribution, the jump time is fitted by exponential distribution and the jump intensity tensor is output; S324, based on the jump probability matrix and the jump intensity tensor, construct a time-varying transfer probability function, calculate the transfer characteristics of different time scales; based on the transfer characteristics, extract and fuse multi-scale transfer information to generate a transfer probability matrix; S325, based on the multi-level feature set, calculating the time series correlation of the feature vector and constructing a covariance estimator; Based on the covariance estimator, the covariance is updated through a sliding window and the time-varying covariance matrix is ​​output; S326, constructing a feature weighting function based on the transition probability matrix and the time-varying covariance matrix, and calculating the importance weight of the state transition; Based on the importance weights, a weighted feature fusion operation is performed to generate a time series feature matrix.

10. A large-scale model for power grid dispatching decision-making based on graph neural network, characterized by: include: at least one processor; as well as, a memory communicatively connected to at least one of the processors; wherein, The memory stores instructions that can be executed by the processor, and the instructions are used to be executed by the processor to implement the power grid dispatching decision method based on graph neural network as described in any one of claims 1 to 9.

Images