Power grid topology optimization method and system based on search sorting
By employing a search-ranking-based power grid topology optimization method, combined with an adaptive dual-tower model and graph attention network, the high computational complexity, dynamic nature, and multi-objective optimization challenges in power grid topology optimization are addressed. This enables rapid and efficient optimization and decision-making for large-scale power grid systems, enhancing the system's adaptability and decision quality.
Patent Information
- Application Number
- CN202411014021.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-26
- Publication Date
- 2024-11-22
- Estimated Expiration
- 2044-07-26
AI Technical Summary
Existing power grid topology optimization methods have high computational complexity when dealing with large-scale, non-convex optimization problems, making it difficult to meet the needs of real-time optimization and adapt to the dynamics and uncertainties of the power grid. In multi-objective optimization, objective trade-offs are difficult to quantify accurately, and efficient screening and evaluation of massive actions are difficult. Constraints are complex and varied, making it difficult to meet various operational constraints while ensuring optimization efficiency. It is also difficult to achieve consistency between short-term and long-term objective decisions.
A power grid topology optimization method based on search ranking is adopted. By constructing an adaptive dual-tower model and graph attention network, and combining it with a time-series graph neural network for multi-objective optimization and decision-making, an adaptive constraint network propagation algorithm and a fuzzy decision forest algorithm are used to handle complex constraints. An adaptive weight adjustment and fuzzy hierarchical analysis network are introduced for multi-objective decision-making. A dynamic dual-tower feature fusion recall model and an adaptive multi-objective optimization method are constructed.
It improves the efficiency and accuracy of power grid topology optimization, enhances the system's adaptability and decision quality, enables rapid optimization of large-scale power grid systems, can dynamically adjust optimization strategies to cope with power grid changes, balance multiple optimization objectives, and meet complex constraints.
Smart Images

Figure CN118539441B_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the field of intelligent dispatching of power grids and discloses a search-sort-based power grid topology optimization method and system. Background Art
[0002] Grid topology optimization is one of the key technologies to ensure the safe, economical and reliable operation of power systems. With the continuous expansion of the scale of power grids, the increasingly complex structure, and the massive access of new energy and flexible loads, grid topology optimization faces unprecedented challenges and opportunities. As an emerging optimization strategy, the grid topology optimization method based on search sorting shows great potential in improving optimization efficiency, adapting to complex grid structures and dynamic operating environments, and has important research significance for improving the overall performance of the grid and enhancing system stability and reliability.
[0003] At present, the research on power grid topology optimization mainly focuses on the following aspects: traditional mathematical programming methods, such as mixed integer linear programming (MILP) and nonlinear programming (NLP); heuristic algorithms, such as genetic algorithms (GA) and particle swarm optimization (PSO); and artificial intelligence-based methods, such as deep reinforcement learning (DRL) and graph neural networks (GNN). These methods have achieved certain results in specific scenarios, but they still have many limitations. Although mathematical programming methods can obtain exact solutions, they have high computational complexity when dealing with large-scale, non-convex optimization problems, and it is difficult to meet real-time optimization requirements. Although heuristic algorithms have a wide range of applications, they are prone to fall into local optimality and are difficult to adjust parameters. Although artificial intelligence-based methods show advantages in dealing with high-dimensional complex problems, they face challenges such as poor model interpretability and limited generalization ability.
[0004] However, there are still several key technical problems to be solved in existing research: First, the dynamics and uncertainty of the power grid topology are difficult to model effectively, which makes it difficult for the optimization results to adapt to the real-time changing power grid status. Second, the trade-offs between the various objectives in the multi-objective optimization problem are difficult to accurately quantify, which affects the practicality of the optimization results. Third, the efficient screening and evaluation of massive possible actions in large-scale power grid systems is still a challenge, which restricts the improvement of optimization efficiency. In addition, the constraints in power grid topology optimization are complex and changeable. How to meet various operating constraints while ensuring optimization efficiency is also a difficult problem. Finally, how to coordinate topology optimization decisions at different time scales to achieve a balance between short-term and long-term goals is also a problem that needs to be solved in current research. These technical difficulties have seriously restricted the practical application and promotion of power grid topology optimization technology, and new optimization methods need to be developed to overcome these challenges. Summary of the invention
[0005] The purpose of the invention is to propose a power grid topology optimization method and system based on search sorting to solve the above-mentioned problems existing in the prior art.
[0006] The technical solution is a power grid topology optimization method based on search sorting, which includes the following steps:
[0007] S1. Acquire power grid topology data, and construct an action space based on the power grid topology data; wherein the power grid topology data includes a node set, an edge set, and an attribute set;
[0008] S2, obtaining real-time grid status data, building an adaptive dual-tower model based on the real-time grid status data and action space, and calculating a candidate action set;
[0009] S3. Based on the candidate action set, a graph attention network and a temporal graph neural network are constructed to obtain a sorted list of candidate actions.
[0010] S4. Based on the sorted candidate action list, multi-objective optimization and decision-making are performed to obtain the optimal action.
[0011] The power grid topology optimization system based on search and sorting includes:
[0012] at least one processor; and,
[0013] a memory communicatively connected to at least one of the processors; wherein,
[0014] 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 topology optimization method based on search sorting described in any of the above technical solutions.
[0015] Beneficial effect: The present invention not only improves the overall effect of power grid topology optimization through innovative algorithm construction and the organic combination of multiple methods, but also enhances the adaptability and decision-making quality of the system, realizes the rapid optimization of large-scale power grid systems, and provides strong support for practical applications. BRIEF DESCRIPTION OF THE DRAWINGS
[0016] Figure 1 It is a flow chart of the present invention.
[0017] Figure 2 This is a flow chart of step S1 of the present invention.
[0018] Figure 3 This is a flow chart of step S2 of the present invention.
[0019] Figure 4 This is a flow chart of step S3 of the present invention.
[0020] Figure 5 This is a flow chart of step S4 of the present invention. DETAILED DESCRIPTION
[0021] like Figure 1 As shown, this application proposes a power grid topology optimization method based on search sorting, and the specific steps are as follows:
[0022] S1. Acquire power grid topology data, and construct an action space based on the power grid topology data; wherein the power grid topology data includes a node set, an edge set, and an attribute set;
[0023] S2, obtaining real-time grid status data, building an adaptive dual-tower model based on the real-time grid status data and action space, and calculating a candidate action set;
[0024] S3. Based on the candidate action set, a graph attention network and a temporal graph neural network are constructed to obtain a sorted list of candidate actions.
[0025] S4. Based on the sorted candidate action list, multi-objective optimization and decision-making are performed to obtain the optimal action.
[0026] In this embodiment, by acquiring real-time grid status data and constructing an adaptive dual-tower model, the topological structure of the grid is dynamically adjusted, and various disturbances and faults in the grid are promptly responded to. This adaptability enables the grid to better maintain stable operation and reduce the risk of large-scale power outages caused by faults; by sorting candidate actions based on the graph attention network and the time-series graph neural network, the most favorable operation for the grid operation can be more accurately identified, so that the economy, reliability and safety of the grid can be comprehensively considered in the multi-objective optimization and decision-making process, thereby achieving the optimal allocation of resources. By using the graph attention network and the time-series graph neural network, the complex data of the grid can be processed and analyzed more efficiently, which not only improves the automation of the grid, but also enhances the self-regulation ability of the grid in the face of complex environments; by constructing and sorting the candidate action list, the necessity of manual intervention is reduced, making the optimization and decision-making process of the grid more efficient and accurate, while also shortening the decision-making time and improving the response speed of the grid. This embodiment dynamically adjusts the optimization strategy according to real-time data and grid status, and enhances the adaptability of the grid in the face of emergencies and changing environments.
[0027] like Figure 2 As shown, according to one aspect of the present application, step S1 is further:
[0028] S11, acquiring power grid topology data and action data affecting the power grid topology; the power grid topology data includes a node set, an edge set and an attribute set; the action data includes an action type, an affected device ID set and an execution parameter set;
[0029] S12, using a graph traversal algorithm to process the power grid topology data, and generating a basic action set based on the action data and the processed power grid topology data;
[0030] S13, based on the basic action set, using a combinatorial optimization algorithm to form an action combination set; based on the action combination set, using a heuristic algorithm to obtain an optimized action space;
[0031] S14, based on the optimized action space and the predefined constraint condition set, sequentially using the constraint propagation algorithm and the decision tree algorithm to obtain the classified action space; based on the classified action space, using the heuristic rule engine to obtain the final optimized action space;
[0032] S15. The finally optimized action space is structured and stored to form a structured data set.
[0033] In this embodiment, the grid topology data and the action data affecting the grid topology are obtained, and the grid topology data is processed using a graph traversal algorithm to generate a basic action set. This basic action set is input into a combinatorial optimization algorithm to generate an action combination set. Finally, the action combination set is input into a heuristic algorithm to obtain an optimized action space.
[0034] The optimized action space and a set of predefined constraints are input into the adaptive constraint network propagation algorithm. The algorithm first constructs a constraint network where nodes represent variables and edges represent constraint relationships. Then the range and constraint strength of each node are initialized. For each constraint, its violation is calculated and the constraint strength is updated. The algorithm selects the strongest constraint for propagation, narrows the range of the relevant node and propagates the impact to the adjacent nodes. This process dynamically adjusts the propagation order, giving priority to high-strength constraints, until the network converges or reaches the iteration limit, and finally generates the constrained action space.
[0035] The constrained action space is input into the fuzzy decision forest algorithm. The algorithm constructs multiple decision trees, each of which uses a random feature subset. Fuzzy splits are introduced at each node, and fuzzy membership functions are defined to allow samples to partially belong to multiple child nodes. During the growth of the tree, the algorithm dynamically adjusts the fuzziness. The results of multiple fuzzy decision trees are integrated using a weighted voting method, and the weights are dynamically adjusted based on the performance of each tree to obtain the classified action space.
[0036] The classified action space and expert knowledge rule set are input into the cognitive graph enhanced rule engine. The engine first constructs an initial cognitive graph, where nodes represent concepts and edges represent relationships, and maps expert rules onto the cognitive graph. For each input action, the relevant nodes are activated on the cognitive graph, and the pulse propagation algorithm is used to simulate the cognitive process and calculate the activation strength of each rule. Then the rule weights are dynamically updated, and each action is re-evaluated based on the updated weights, and Top-K filtering is applied to obtain the final optimized action space.
[0037] The optimized action space is structured and stored, specifically: a unique identifier is assigned to each action to form an action index table. The specific parameters of each action are stored to form an action parameter matrix. Based on the action parameter matrix, an association graph is formed to represent the dependency and mutual exclusion relationship between actions. These three components together constitute the final structured data set.
[0038] In a further embodiment, power grid topology data and action data that may affect the power grid topology are obtained, wherein the power grid topology data includes a node set, an edge set, and an attribute set, and the action data includes an action type, an affected device ID set, and an execution parameter set;
[0039] The graph traversal algorithm is used to process the power grid topology data to generate a basic action set. The basic action set is then input into the combinatorial optimization algorithm to generate an action combination set. The action combination set is then input into the heuristic algorithm to finally obtain the optimized action space.
[0040] The optimized action space and the set of predefined constraints are input into the adaptive constraint network propagation algorithm (ACNP) to generate the constrained action space; then the constrained action space is input into the fuzzy decision forest algorithm (FDF) to obtain the classified action space; then the classified action space and the expert knowledge rule set are input into the cognitive graph enhanced rule engine (CGRE) to finally obtain the optimized action space.
[0041] When using the ACNP algorithm, a constraint network is constructed, where nodes represent variables and edges represent constraint relationships. Initialize the value range D(i) and constraint strength CS(i, j) of each node. For each constraint C(i, j): calculate the constraint violation degree CV(i, j) = |D(i) ∩ D(j)| / |D(i) ∪ D(j)|, update the constraint strength: CS(i, j) = CS(i, j) * (1 + η * CV(i, j)), where η is the learning rate; select the constraint with the highest strength for propagation, reduce the value range of the relevant nodes, and propagate the impact to the adjacent nodes; dynamically adjust the propagation order and give priority to high-strength constraints. Repeat until the network converges or reaches the iteration limit, and output the action space that satisfies the constraints.
[0042] When using the FDF algorithm, multiple decision trees are constructed, and each tree uses a random feature subset. Fuzzy splitting is introduced at each node, and a fuzzy membership function μ(x) is constructed. Samples can partially belong to multiple child nodes. During the tree growth process, the fuzziness θ is dynamically adjusted: θ = θ0 * (1 - depth / max depth ) ρ , where ρ is the attenuation factor, depth represents the depth of the current node in the decision tree, and max depth Represents the length of the longest path that a decision tree can generate, and θ0 represents the initial fuzziness. The results of multiple fuzzy decision trees are integrated, and the weighted voting method is used to dynamically adjust the weights based on the performance of each tree to obtain the classified action space.
[0043] When using the CGRE algorithm, an initial cognitive graph is constructed, where nodes represent concepts and edges represent relationships. Expert rules are mapped to the initial cognitive graph. For each input action in the action space, relevant nodes are activated on the initial cognitive graph CG, and the pulse propagation algorithm is used to simulate the cognitive process, and the activation strength AS(r) of each rule is calculated; the rule weights are dynamically updated: w(r) =w(r) * (1 + ξ * AS(r)), where ξ is the update rate; based on the updated weights, each action is re-evaluated, and Top-K screening is applied to obtain the optimized action space.
[0044] The optimized action space is structured and stored. A unique identifier is assigned to each action to form an action index table I. The specific parameters of each action are stored to form an action parameter matrix M. The dependency and mutual exclusion relationships between actions are represented to form an association graph G'. Finally, a structured data set D = {I, M, G'} is obtained.
[0045] In another specific embodiment, the system collects and organizes the complete topological structure information of the current power grid to generate a power grid topology data set T. The power grid topology data set T includes a node set N: including power stations, substations, load points, etc.; an edge set E: representing the connection relationship between nodes; and an attribute set A: recording various attributes of nodes and edges (such as capacity, impedance, etc.).
[0046] The system identifies and records all operations that may affect the topology of the power grid and generates a set of original actions O. Each action o∈O contains: action type C: such as switching lines, adjusting transformer taps, etc.; affected device ID set D; execution parameters P: such as switch status, adjustment amplitude, etc.
[0047] Based on the power grid topology dataset T, a graph traversal algorithm is used to perform a depth-first search (DFS) or a breadth-first search (BFS) to obtain a basic action set B. By traversing the entire power grid topology, possible basic actions (such as switch operation, adjusting transformer taps, etc.) are generated for each node and edge. Based on the basic action set B, a combinatorial optimization algorithm is used to obtain an action combination set C. By generating an effective combination of basic actions, the mutual influence and constraints between actions are taken into account. Based on the action combination set C, a simulated annealing algorithm or a genetic algorithm is used to obtain an optimized action space S. By using a heuristic algorithm to preliminarily optimize the action combination, an action sequence with potential optimization effect is generated.
[0048] The optimized action space S is further screened and optimized. Based on the optimized action space S and the predefined constraint condition set, the AC-3 algorithm (arc consistency algorithm) is used to obtain the constrained action space S1, so that actions that do not meet the basic constraints can be quickly eliminated. Based on the constrained action space S1, the decision tree algorithm is used to obtain the classified action space S2, so that actions can be classified and further screened based on safety, feasibility and other conditions. Based on the classified action space S2 and the expert knowledge rule set R, the Rete algorithm is used to obtain the optimized action space S', so that the rule set constructed by expert knowledge can be applied to finally evaluate and screen the actions.
[0049] The optimized action space S' is stored as structured data D, including an action index table I: assigning a unique identifier to each action; an action parameter matrix M: storing the specific parameters of each action; and an association graph G: representing the dependency and mutual exclusion relationships between actions.
[0050] This embodiment improves the efficiency and accuracy of power grid topology optimization through data preprocessing and action space construction methods. A comprehensive data foundation is established by obtaining power grid topology data and action data that may affect the power grid topology. This comprehensive data collection ensures the integrity and reliability of the subsequent optimization process. The power grid topology data is processed by a graph traversal algorithm to generate a basic action set, and then the action space is gradually optimized by combining optimization algorithms and heuristic algorithms. This multi-level action space construction method not only considers a single action, but also considers the effect of action combinations, expanding the range of possible solutions. The adaptive constraint network propagation algorithm (ACNP) is introduced to process the optimized action space, so that the constraint strength and propagation order can be dynamically adjusted, complex constraints can be effectively processed, and the generated solution can be ensured to meet the actual operation requirements. The fuzzy decision forest algorithm (FDF) is used to classify the constrained action space, and the flexibility and accuracy of the classification are improved by introducing fuzzy splitting and dynamically adjusting the fuzziness. The cognitive graph enhanced rule engine (CGRE) is used for the final optimization. By simulating the human cognitive process and combining expert knowledge, a more intelligent optimization result that is more in line with actual operation experience is achieved.
[0051] This embodiment achieves efficient construction and optimization of the action space, improves the starting point of the subsequent optimization process, and lays a solid foundation for the entire power grid topology optimization process. In addition, the final structured storage construction, including the construction of the action index table, action parameter matrix, and association graph, facilitates subsequent rapid retrieval and analysis, and further improves the operating efficiency of the entire system.
[0052] In another embodiment, a multi-level adaptive graph traversal algorithm (MLAGT) can also be used to process the power grid topology data T to generate a basic action set B; then the basic action set B is input into a dynamic weight combination optimization algorithm (DWCO) to generate an action combination set C; and then the action combination set C is input into a self-organizing critical heuristic algorithm (SOCH) to finally obtain an optimized action space S. Specifically, multiple traversal levels L1, L2, ..., Ln are initialized, each level corresponds to a different traversal granularity; traversal starts from the coarsest granularity level L1 to identify key nodes and edges; for each key node, a finer granularity traversal is performed in the next level L2; during the traversal process, the importance weight W(i) of each node is dynamically adjusted based on its connectivity and load conditions; an adaptive step size λ(t) is used to switch between different levels, λ(t) = λ0 / (1 + αt), where t is the number of iterations and α is the decay rate; repeat until the finest granularity level Ln is reached or the stop condition is met; and a basic action set B is generated based on the traversal results and node weights.
[0053] Initialize the action combination set C0, set the initial temperature T0 and the cooling coefficient β; for each action ai Assign initial weight w i ; In each iteration, generate a new action combination C' based on the current temperature T and action weight w; Calculate the objective functions f(C') and f(C); If f(C') < f(C), accept the new combination; otherwise accept with probability p = exp((f(C)-f(C')) / T); Update the action weight: w i = w i * (1 + γ *Δf), where γ is the learning rate and Δf is the improvement of the objective function; reduce the temperature: T = T * β; repeat until the stopping condition is reached and output the optimal action combination set C.
[0054] Initialize the action sequence S0, set the initial energy E0 and critical exponent τ; define the energy function E(S) to evaluate the quality of the action sequence; in each iteration, make a small perturbation to the current sequence S to generate a new sequence S'; calculate the energy difference ΔE = E(S') - E(S); if ΔE < 0, accept the new sequence; otherwise, accept it with probability p = (E(S) / E(S')) τ Accept; update the critical index: τ = τ0 * log(1 + t), where t is the number of iterations; dynamically adjust the disturbance amplitude to keep the system in a critical state; repeat until the stopping condition is reached; output the optimized action sequence S.
[0055] like Figure 3 As shown, according to one aspect of the present application, step S2 is further:
[0056] S21, acquiring real-time power grid status data, and using a multi-scale time series feature extraction algorithm based on the real-time power grid status data to obtain a dynamic feature vector;
[0057] S22, based on the power grid topology data, using the adaptive spectral clustering enhancement algorithm to obtain a static feature vector;
[0058] S23, based on the dynamic feature vector and the static feature vector, adopting an adaptive weighted fusion algorithm to obtain a fused feature vector;
[0059] S24, constructing an adaptive dual-tower model, inputting the fused feature vector into the state tower in the dual-tower model to obtain a state embedding vector; inputting the finally optimized action space into the action tower in the dual-tower model to obtain an action embedding vector set;
[0060] S25. Based on the state embedding vector and the action embedding vector set, an adaptive similarity calculation and a dynamic threshold selection algorithm are used to obtain a candidate action set.
[0061] Step S21 is further as follows:
[0062] S211, acquiring real-time power grid status data, and extracting time series from the real-time power grid status data;
[0063] S212, decomposing the time series into predetermined time scales, using an adaptive window sliding algorithm for each time scale, and calculating the size of each window;
[0064] S213, based on the real-time power grid status data, calculating statistical features in each window, and using wavelet transform to extract frequency domain features in the statistical features; wherein the statistical features include mean, variance and kurtosis;
[0065] S214, constructing a feature importance evaluation index, and updating the weight of the frequency domain feature based on the feature importance evaluation index;
[0066] S215: concatenate the frequency domain features after the weights are updated to form a final dynamic feature vector.
[0067] In a specific embodiment, real-time power grid status data is obtained and processed using a multi-scale time series feature extraction algorithm. The algorithm decomposes the input time series data into multiple time scales. For each time scale, an adaptive window sliding algorithm is applied, and the window size is dynamically adjusted with the time scale. Statistical features such as mean, variance, and kurtosis are calculated in each window, and frequency domain features are extracted using wavelet transform. Importance weights are applied to each feature, and the initial value is uniformly distributed. A feature importance evaluation index is constructed based on its predictive ability for the system state. The feature weights are updated regularly, and the weighted features of all scales are spliced to form the final dynamic feature vector.
[0068] The grid topology dataset is processed by adaptive spectral clustering enhancement algorithm. The initial similarity matrix is constructed based on the physical connection and electrical characteristics between nodes. The adaptive kernel function is introduced to calculate the Laplace matrix. The generalized eigenvalue problem is solved to obtain the first k eigenvectors. The self-organizing map is applied in the feature space to form the initial clustering. The concept of fuzzy clustering is introduced to allow nodes to partially belong to multiple clusters. The clustering effectiveness index is defined to periodically evaluate the clustering quality. If the index decreases, the k value and kernel function parameters are dynamically adjusted. This process is repeated until convergence or the iteration limit is reached. Finally, based on the final clustering result, the static eigenvector is extracted.
[0069] The dynamic feature vector and the static feature vector are input into the dynamic attention fusion network. A dual-channel neural network is constructed to process dynamic and static features respectively. In each channel, a multi-head self-attention mechanism is used to capture the internal relationship of the features, and residual connections and layer normalization are applied. Then a cross-channel attention module is constructed to calculate the attention weight matrix between dynamic and static features. A gating mechanism is used to dynamically adjust the importance of the two features. Finally, the features are fused, nonlinear transformations and residual connections are applied, and the final fused feature vector is output.
[0070] Construct an adaptive dual-tower model. Input the fused feature vector into the state tower composed of the dynamic structure multi-layer perceptron to obtain the state embedding vector. At the same time, input the optimized action space into the action tower composed of hierarchical local sensitive hashing to obtain the action embedding vector set. The dynamic structure multi-layer perceptron dynamically adjusts the network structure by defining the neuron importance index, removes unimportant neurons, and adds new connections or layers between important layers. Hierarchical local sensitive hashing divides the action space into multiple levels, each representing a different granularity. A set of hash functions is used to map the action to multiple buckets, and an inter-layer index structure is constructed to allow fast switching between different granularities.
[0071] Using the state embedding vector and action embedding vector set, an adaptive similarity calculation and dynamic threshold selection algorithm is adopted. Initialize multiple similarity metrics, such as cosine similarity, Euclidean distance, etc., and assign initial weights. Calculate weighted similarity, and then use the kernel density estimation method to estimate the similarity distribution. Dynamically determine the threshold and select actions with similarity greater than the threshold as candidates. If the size of the candidate set exceeds the predetermined value, use the diversity sampling algorithm to filter it. Finally, update the similarity weight and output the final Top-K candidate action set.
[0072] In a further embodiment, real-time grid status data is obtained and processed using a multi-scale time series feature extraction algorithm (MSTFE) to obtain a dynamic feature vector, specifically:
[0073] Decompose the input time series R into multiple time scales {R1, R2, ..., R n}; For each scale R i , apply the adaptive window sliding algorithm, the window size W'(i) = W'0 * (1 + β' i ), where W'0 represents the initial window size, β' irepresents the adjustment factor related to the i-th time scale; calculates statistical features (mean, variance, kurtosis, etc.) within each window; uses wavelet transform to extract frequency domain features; applies importance weight ω(i) to each feature, with the initial value uniformly distributed; constructs the feature importance evaluation index I(f) based on its predictive ability for the system state, and regularly updates the feature weight: ω(i) = ω(i)* (1 + γ * I(f)), where γ is the learning rate; concatenates the weighted features of all scales to form the final dynamic feature vector D.
[0074] The power grid topology dataset is processed by the adaptive spectral clustering enhancement algorithm (ASCE) to obtain the static feature vector F, which is:
[0075] Construct the initial similarity matrix S″ based on the physical connection and electrical characteristics between nodes; introduce the adaptive kernel function K(x, y) = exp(-d(x, y) 2 / (2σ(x)σ(y))), where σ(x) and σ(y) are local scale parameters, and d(x, y) represents the distance between a node and node y; calculate the Laplace matrix L = D" - S", where D" is the degree matrix; solve the generalized eigenvalue problem Lv = λ'D"v, and obtain the first k eigenvectors, where λ' represents the eigenvalue in the generalized eigenvalue problem and v represents the eigenvector; apply self-organizing map (SOM) in the feature space to form initial clustering; introduce the concept of fuzzy clustering, allow nodes to partially belong to multiple clusters, define the clustering validity index CVI, periodically evaluate the clustering quality, and if CVI decreases, dynamically adjust the k value and kernel function parameters, and repeat until convergence or the iteration limit is reached; based on the final clustering result, extract the static eigenvector F.
[0076] The dynamic feature vector D and the static feature vector F are input into the dynamic attention fusion network (DAFN) to obtain the fused feature vector H, which is specifically:
[0077] A dual-channel neural network is constructed to process the dynamic feature vector D and the static feature vector F respectively; in each channel, a multi-head self-attention mechanism is used to capture the internal relationship of features, and residual connections and layer normalization are applied; a cross-channel attention module is constructed to calculate the attention weight matrix A between the dynamic feature vector D and the static feature vector F.
[0078] A(i, j) = softmax(Q(D i )K(F j ) T / sqrt(d));
[0079] Where Q and K are linear transformations, and d represents the dimension of the feature vector;
[0080] A gating mechanism G is used to dynamically adjust the importance of two features:
[0081] G = σ(W g [D; F] + b g );
[0082] Where σ is the sigmoid function, b g Represents the bias term in the gating mechanism.
[0083] Fusion features: H = G * (AD) + (1-G) * (A T F); apply nonlinear transformation and residual connection: H = H + FFN(H), and output the final fused feature vector H.
[0084] Construct an adaptive dual-tower model, input the fused feature vector H into the state tower composed of the dynamic structure multi-layer perceptron (DSMLP), and obtain the state embedding vector S emb ; At the same time, the optimized action space S' is input into the action tower composed of hierarchical local sensitive hashing (HLSH) to obtain the action embedding vector set A emb , specifically:
[0085] Initialize an MLP structure with variable width and depth; define a neuron importance index NI based on its activation frequency and gradient; in each training cycle, forward propagate to calculate the loss, back propagate to update the parameters, and evaluate the neuron importance index NI of each neuron; perform structural optimization regularly, remove neurons with neuron importance index NI below the threshold, and add new connections or layers between layers with high NI; use dynamic Dropout, with the Dropout rate inversely proportional to NI, and output the final state embedding vector S emb .
[0086] The action space S' is divided into multiple levels {L1, L2, ..., Ln}, each level represents a different granularity; at each level Li, a set of hash functions h is constructed i,j (x) = sign(a i,j T x + b i,j ), where a i,j represents the weight vector in the hash function, b i,j Represents the bias term in the hash function, mapping actions to multiple buckets; constructs an inter-layer index structure to allow fast switching between different granularities; for the query vector q, starts searching for the nearest neighbor from the top layer L1, and dynamically decides whether to drill down to the next layer based on the matching quality of the current layer; during the search process, uses local sensitive forest (LSF) to improve accuracy, builds multiple LSH trees, integrates the results of multiple trees, and improves the recall rate; outputs the action embedding vector set A emb .
[0087] Using the state embedding vector S emb and action embedding vector set A emb , using the adaptive similarity calculation and dynamic threshold selection algorithm (ASCDTS), we get the Top-K candidate action set C, specifically:
[0088] Initialize multiple similarity metrics {sim1, sim2, ..., sim m}, including cosine similarity, Euclidean distance, etc.; for each similarity measure sim i , assign initial weight w i ; Calculate weighted similarity:
[0089] Sim W (S emb , a) = Σ(w i * sim i (S emb , a)), a ∈ A emb .
[0090] Use the kernel density estimation (KDE) method to estimate the similarity distribution P(Sim W ), dynamically determine the threshold θ: initialize θ0 = μ + σ, where μ and σ are the mean and standard deviation of the weighted similarity; in each iteration, adjust θ according to the results of the previous round:
[0091] θ t = θ t-1 + λ(K - |C t-1 |);
[0092] Where λ is the learning rate, C t-1 is the action set selected in the previous round, and K represents the target size of the candidate action set.
[0093] Select actions with similarity greater than θ as candidates; if the candidate set size |C| > K, use the diversity sampling algorithm to screen, greedily select the first action, and iteratively select the action with the smallest similarity to the selected action set. Update the similarity weight:
[0094] w i = w i * (1 + η * performance(C i ));
[0095] η is the update rate, performance(C i ) is to use sim i The performance scores of the selected action sets are calculated, and the final Top-K candidate action set C is output.
[0096] In another specific embodiment, a time series algorithm is used to process real-time grid state data R to capture the time series characteristics of grid load, voltage and other parameters; a graph algorithm is used to process the grid topology data set T to extract the static characteristics of the network structure; the dynamic characteristics and static characteristics are fused to generate a comprehensive feature representation, that is, a fused feature vector H. A dual-tower model is constructed to process state characteristics and action characteristics respectively. The fused feature vector H is input into the state tower and a multi-layer perceptron (MLP) is performed to obtain the state embedding vector S emb ; Input the optimized action space S' into the action tower and perform local sensitive hashing (LSH) to obtain the action embedding vector set A emb ; Based on the state embedding vector S emb and action embedding vector set A emb , using cosine similarity and Top-K selection algorithm, calculate the similarity between the state embedding vector and the action embedding vector, recall the most relevant candidate actions, and obtain the Top-K candidate action set C. This combines dynamic and static features to fully capture the real-time state and topological structure of the power grid. At the same time, an adaptive weighted fusion method is used to flexibly adjust the importance of different features; a dual-tower model structure is used to improve the recall efficiency in large-scale action space; and local sensitive hashing is used to process action features, further improving the computational efficiency of the model.
[0097] This embodiment greatly improves the efficiency and accuracy of power grid state feature extraction and action recall through a dynamic dual-tower feature fusion recall model. A multi-scale time series feature extraction algorithm is used to process real-time power grid state data. This algorithm effectively captures features of different time scales by decomposing the time series into multiple scales and applying an adaptive window sliding algorithm on each scale. At the same time, by dynamically adjusting the feature weights, it is ensured that the extracted features can accurately reflect the key information of the current power grid state. This method improves the comprehensiveness and adaptability of feature extraction and provides a more reliable data basis for subsequent decision-making. The application of the adaptive spectral clustering enhancement algorithm improves the analysis ability of the power grid topology structure by introducing adaptive kernel functions and fuzzy clustering concepts. This method can better capture the complex relationships in the network and provide a more accurate static feature representation. The use of the dynamic attention fusion network realizes the effective fusion of dynamic features and static features through a dual-channel neural network and a cross-channel attention mechanism. This fusion method not only considers the relationship within the features, but also captures the interaction between different types of features through the attention mechanism, thereby generating a more comprehensive and accurate fusion feature representation. The construction of the adaptive dual-tower model, including the application of dynamic structure multi-layer perceptron and hierarchical local sensitive hashing, further improves the flexibility and efficiency of the model. The dynamic structure multi-layer perceptron can dynamically adjust the network structure according to the importance of features, while the hierarchical local sensitive hashing realizes efficient indexing and retrieval of large-scale action spaces. The use of adaptive similarity calculation and dynamic threshold selection algorithm ensures the accuracy and diversity of candidate action selection through the combination of multiple similarity metrics and dynamic threshold adjustment. It not only improves the quality of recall, but also maintains the adaptability of the system through dynamic adjustment.
[0098] This embodiment achieves accurate perception of the power grid status and efficient action recall, improves the basis for subsequent decision-making, and also reduces the computational complexity, making it possible to optimize decisions in real time.
[0099] like Figure 4 As shown, according to one aspect of the present application, step S3 is further:
[0100] S31, acquiring historical power grid status data, and constructing a dynamic space-time graph based on the historical power grid status data and power grid topology data;
[0101] S32. Based on the dynamic spatiotemporal graph and state embedding vector, a multi-scale dynamic graph attention network is constructed to obtain attention-weighted node representation;
[0102] S33. Based on the attention-weighted node representation and candidate action set, an adaptive spatiotemporal graph convolutional network is constructed to obtain the spatiotemporal representation of action-state pairs.
[0103] S34. Based on the spatiotemporal representation of action-state pairs, a hierarchical multi-criteria decision network is constructed to generate the final scores of candidate actions and a ranked list of candidate actions.
[0104] Step S32 is further as follows:
[0105] S321. Construct a multi-scale dynamic graph attention network based on dynamic spatiotemporal graph and state embedding vector;
[0106] S322. Based on the state embedding vector, calculate the query, key, and value matrices of each attention head in the multi-scale dynamic graph attention network;
[0107] S323, based on the query, key, and value matrix, calculate an attention score, introduce a time decay factor to modify the attention score, and obtain a modified attention score;
[0108] S324, based on the modified attention score and value matrix, calculate the output of each attention head, use the gating mechanism to fuse the output of each attention head, and obtain a fusion result;
[0109] S325. Based on the fusion result and the state embedding vector, residual connection and layer normalization are used to calculate the attention-weighted node representation.
[0110] In a specific embodiment, a power grid topology data set and historical power grid state data are obtained and processed using a spatiotemporal adaptive graph construction algorithm. The algorithm initializes a static graph structure based on the power grid topology data. The historical data is divided into time windows, and the window size is dynamically adjusted. In each time window, the dynamic association strength between nodes is calculated and the edge weight is updated. Edges are dynamically added or removed according to the weights. An adaptive threshold is used to control the change of the graph structure. Finally, the edges of the time dimension are constructed to connect the same nodes at different time steps to obtain a dynamic spatiotemporal graph.
[0111] The dynamic spatiotemporal graph and state embedding vectors are fed into a multi-scale dynamic graph attention network. The network defines multiple attention heads, each focusing on a different feature subspace. For each attention head, the query, key, and value matrices are calculated and the attention score is computed. A time decay factor is introduced to modify the attention calculation. The output is calculated and the multi-head results are fused using a gating mechanism. Finally, residual connections and layer normalization are applied to obtain attention-weighted node representations.
[0112] Using the attention-weighted node representation and candidate action set, an adaptive spatiotemporal graph convolutional network is used for processing. The network constructs an adaptive graph convolution filter and introduces a dynamic adjacency matrix. A multi-layer network structure is constructed, including spatial convolutional layers, temporal convolutional layers, and combination layers. The outputs of different layers are aggregated using the attention mechanism. For each candidate action, an action embedding is generated, and the action-state pair representation is calculated to obtain the spatiotemporal representation of the action-state pair.
[0113] The spatiotemporal representation of the action-state pairs is input into a hierarchical multi-criteria decision-making network. The network first defines multiple evaluation criteria, such as safety, economy, reliability, etc. A hierarchical neural network structure is constructed, including the bottom sub-network, the middle layer, and the top layer. The attention mechanism is applied in each sub-network to dynamically adjust the feature importance. A soft hierarchical decomposition is used to allow part of the evaluation results to flow to non-direct superior nodes. Uncertainty modeling is introduced, and probability distribution is used to represent the evaluation results. The gradient boosting method is applied to optimize the network parameters. Finally, the final score and ranking list of candidate actions are generated.
[0114] In a further embodiment, a power grid topology data set T and historical power grid status data H are obtained and processed using a spatiotemporal adaptive graph building algorithm (STAGBA) to obtain a dynamic spatiotemporal graph G ts , specifically:
[0115] Based on the power grid topology data T, initialize the static graph structure G s . Divide the historical data H into time windows, and adjust the window size w dynamically; in each time window t, calculate the dynamic correlation strength R(i, j, t) between nodes, and update the edge weight: W(i, j, t) = α * W(i, j, t-1) + (1-α) * R(i, j, t); if W(i, j, t) > θ, add a temporary edge; if W(i, j, t) < φ, remove the existing edge; use an adaptive threshold:
[0116] θ t = θ0 * (1 + β * var(R));
[0117] φ t = φ0 * (1 - β * var(R));
[0118] Where α represents the smoothing factor, θ represents the threshold for adding temporary edges, φ represents the threshold for removing existing edges, and β represents the adaptive threshold adjustment factor.
[0119] Construct edges in the time dimension, connect the same nodes at different time steps, and output a dynamic spatiotemporal graph G ts .
[0120] The dynamic space-time graph G ts and the state embedding vector S emb Input the multi-scale dynamic graph attention network (MDGAN) to get the attention-weighted node representation N att , specifically:
[0121] Define multiple attention heads, each focusing on a different feature subspace; for each attention head h, calculate the query, key, and value matrices:
[0122] Q h =WQ h * S emb ;
[0123] K h =WK h * S emb ;
[0124] V h =WV h * S emb ;
[0125] Where Q h represents the query matrix of attention head h, K h represents the key matrix of the attention head h, V h A matrix representing the values of the attention head h;
[0126] Calculate attention score: A h = softmax((Q h * K h T ) / sqrt(d k + M)), M is the adjacency matrix mask, d k is the dimension of the key vector.
[0127] Introduce the time decay factor: D(t) = exp(-λ *Δt), λ is the decay rate, Δt represents the time difference; modify the attention calculation: A h ' = A h * D(t); Calculation output: O h = A h ' * V h ; Use gating mechanism to fuse multiple results:
[0128] g h = sigmoid(W gh * [O h ; S emb ]) O = Σ(g h * O h );
[0129] Apply residual connections and layer normalization: N att = LayerNorm(O + S emb ).
[0130] Using attention-weighted node representation N attAnd the candidate action set C is processed by the adaptive spatiotemporal graph convolutional network (ASTGCN) to obtain the spatiotemporal representation TS of the action-state pair, specifically:
[0131] Construct an adaptive graph convolution filter: f(X, A) = σ(D -1 / 2 AD -1 / 2 XW), where A represents the adjacency matrix, σ represents the activation function, D represents the degree matrix, X represents the node feature matrix, and W represents the weight matrix; the dynamic adjacency matrix A is introduced t , updated in real time according to the current network status; building a multi-layer ASTGCN structure, spatial convolution layer: GC s (X t ) = f(X t , A t ); Temporal convolution layer: GC t (X) = Conv1D(X), where D(X) represents the result of normalizing the node feature matrix X; combination layer: H l = GC s (GC t (H l -1)); Use attention mechanism to aggregate the outputs of different layers: α l = softmax(v T tanh(WH l + b)) O = Σ(α l * H l ), where b represents the bias term; for each candidate action c ∈ C, generate an action embedding e c ; Compute action-state pair representation: TS c = FC([O; e c ]), output the spatiotemporal representation TS of the action-state pair.
[0132] The spatiotemporal representation TS of the action-state pair is input into the hierarchical multi-criteria decision network (HMCDN) to obtain the sorted candidate action list R C , specifically:
[0133] Define multiple evaluation criteria {C1, C2, ..., C m}, such as safety, economy, reliability, etc.; construct a hierarchical neural network structure, and the underlying sub-network targets each criterion C iThe evaluation is performed; the middle layer integrates the evaluation results of relevant criteria; the top layer generates the final comprehensive score. The attention mechanism is applied in each sub-network to dynamically adjust the feature importance; soft hierarchical decomposition is used: the evaluation results are allowed to flow partially to non-direct superior nodes. Uncertainty modeling is introduced, and the evaluation results are represented by probability distribution rather than a single score, and uncertainty propagation is handled through Monte Carlo sampling. The gradient boosting method is applied to optimize the network parameters and generate the final score and ranking list R of the candidate actions. C .
[0134] In another specific embodiment, the power grid topology data set is combined with the historical power grid state data to construct a time-series graph representation. The time-series graph and state embedding vector are based on using a graph attention network to capture the importance relationship between nodes and obtain an attention-weighted node representation. Based on the attention-weighted node representation and the candidate action set, a time-series graph neural network is used to model the dynamic changes of the power grid state and obtain a time-series representation of the action-state pair. Based on the time-series representation, a multi-layer perceptron (MLP) and a sorting algorithm are used to evaluate and sort the candidate actions to obtain a sorted candidate action list.
[0135] This embodiment improves the accuracy and efficiency of action evaluation in power grid topology optimization through graph attention and sequential graph neural network action evaluation method. The application of spatiotemporal adaptive graph construction algorithm realizes accurate modeling of power grid topology structure changing over time by dynamically adjusting time windows and edge weights. It not only captures the static topology structure, but also takes into account the dynamic changes in the network, providing a more comprehensive and accurate graph structure representation for subsequent analysis. The introduction of multi-scale dynamic graph attention network realizes effective attention to different feature subspaces and time scales through multiple attention heads and time decay factors. In this way, the complex spatiotemporal dependencies in the power grid can be captured, and the accuracy and expression ability of node representation can be improved. The use of adaptive spatiotemporal graph convolution network realizes effective processing of spatiotemporal graph data by constructing adaptive graph convolution filters and dynamic adjacency matrices. The construction of multi-layer network structure, including spatial convolution layer, temporal convolution layer and combination layer, enables the model to capture the characteristics of spatial and temporal dimensions at the same time, thereby generating a more comprehensive action-state pair representation. Finally, the application of hierarchical multi-criteria decision network realizes comprehensive evaluation of candidate actions by defining multiple evaluation criteria and constructing hierarchical neural network structure. The introduction of soft hierarchical decomposition and uncertainty modeling further improves the flexibility and robustness of the evaluation. Not only multiple decision criteria are considered, but also the evaluation results are represented by probability distribution, which better handles the uncertainty in the decision-making process.
[0136] This embodiment achieves accurate evaluation of power grid actions. It not only takes into account the static structure and dynamic changes of the power grid, but also ensures the comprehensiveness and reliability of the evaluation results through a multi-level, multi-criteria evaluation method. This method improves the accuracy of action evaluation and provides a more reliable basis for subsequent optimization decisions. At the same time, it also achieves rapid evaluation of large-scale power grid systems through an efficient graph neural network structure, making real-time decision-making possible.
[0137] like Figure 5 As shown, according to one aspect of the present application, step S4 is further:
[0138] S41, based on a predefined set of optimization objectives, a dynamic weight adaptive fusion algorithm is used to obtain an adaptive multi-objective evaluation function; wherein the set of optimization objectives includes load balancing, network stability and energy efficiency;
[0139] S42, based on the preset power grid operation constraints and the adaptive multi-objective evaluation function, a dynamic constraint processing and an adaptive penalty algorithm are used to obtain an adaptive multi-objective evaluation function with constraints;
[0140] S43, based on the constrained adaptive multi-objective evaluation function and the ranked candidate action list, a co-evolutionary multi-population particle swarm optimization algorithm is used to obtain the Pareto optimal solution set;
[0141] S44. Based on the Pareto optimal solution set, a fuzzy hierarchical analysis network is constructed to obtain the action with the highest comprehensive score as the optimal action.
[0142] Step S41 is further as follows:
[0143] S411, obtaining an optimization target set, and initializing a target weight vector of the optimization target set;
[0144] S412, obtaining historical performance data and current system status data, and constructing an importance index for each target based on the historical performance data and the current system status data;
[0145] S413, based on the importance index, updating the target weight vector to obtain an updated weight, and normalizing the updated weight using a Softmax function to obtain a normalized weight;
[0146] S414, constructing a fuzzy membership function, and obtaining an adaptive multi-objective evaluation function based on the fuzzy membership function and the normalized weights.
[0147] Step S42 is further as follows:
[0148] S421, obtaining preset power grid operation constraint conditions, and classifying the preset power grid operation constraint conditions into hard constraints and soft constraints;
[0149] S422, based on the hard constraints, using feasibility rules to quickly screen the adaptive multi-objective evaluation function to obtain an adaptive multi-objective evaluation function that satisfies the hard constraints;
[0150] S423. Based on soft constraints, construct a default degree function and an adaptive penalty factor;
[0151] S424, constructing a constrained adaptive multi-objective evaluation function based on the adaptive multi-objective evaluation function that satisfies the hard constraints, the breach degree function and the adaptive penalty factor;
[0152] S425. Collect historical data on violations of preset grid operation constraints, increase penalty weights for frequently violated preset grid operation constraints, and obtain weighted preset grid operation constraints; based on the weighted preset grid operation constraints, dynamically adjust the classification of hard constraints and soft constraints.
[0153] In a specific embodiment, a set of optimization targets is obtained and processed using a dynamic weight adaptive fusion algorithm. A target weight vector is initialized. The importance index of each target is defined based on historical performance and current system status. An adaptive weight update rule is constructed, and the updated weight is normalized by applying a Softmax function. A fuzzy membership function is constructed to indicate the degree to which the solution satisfies the target. An adaptive multi-target evaluation function is defined, and the weight update strategy is periodically evaluated and adjusted. Finally, an adaptive multi-target evaluation function is obtained.
[0154] Input the grid operation constraints and adaptive multi-objective evaluation function into the dynamic constraint processing and adaptive penalty algorithm. Classify the constraints into hard constraints and soft constraints. For hard constraints, use feasibility rules to quickly screen solutions that meet the hard constraints, and construct repair operations to map infeasible solutions to the feasible domain boundary. For soft constraints, define the breach degree function and construct an adaptive penalty factor. Construct an evaluation function with constraints, dynamically adjust the penalty strategy, and increase the penalty weight for frequently violated constraints. Re-evaluate the constraint classification regularly and switch between hard and soft constraints when necessary. Finally, obtain a constrained adaptive multi-objective evaluation function.
[0155] The co-evolutionary multi-population particle swarm optimization algorithm is used to process the problem using a constrained adaptive multi-objective evaluation function and a sorted list of candidate actions. Multiple subpopulations are initialized, each focusing on a specific target or search area. In each subpopulation, an improved speed update rule is used to apply adaptive inertia weights and acceleration coefficients. A co-evolutionary mechanism is introduced to regularly exchange elite individuals between subpopulations, and a competition-cooperation strategy is used to evaluate individual fitness. A diversity preservation strategy is constructed, using crowding distance to maintain the diversity of the Pareto frontier and introducing small probability mutations to jump out of the local optimum. The number and size of subpopulations are dynamically adjusted, similar subpopulations are merged, and subpopulations with strong exploratory potential are split. An archiving strategy is applied to maintain the global non-dominated solution set. The search space is reconstructed regularly to focus on potential areas. Finally, the Pareto optimal solution set is obtained.
[0156] Input the Pareto optimal solution set into the fuzzy hierarchical analysis network. Construct a decision hierarchy, including the target layer, the criterion layer, and the solution layer. In each layer, establish a fuzzy comparison matrix. Use the improved eigenvector method to calculate the weights, construct the fuzzy consistency matrix, solve the fuzzy characteristic equation, and reduce the fuzzy weights. Introduce a neural network to process the relationship between layers, use the BP network to learn the mapping relationship between layers, and apply dropout to prevent overfitting. Construct an adaptive membership function and use Choquet fuzzy integral to aggregate multi-criteria evaluation. Apply sensitivity analysis to evaluate the robustness of the results. Finally, output the action with the highest comprehensive score as the optimal action.
[0157] In a further embodiment, an optimization target set O is obtained and processed using a dynamic weight adaptive fusion algorithm (DWAFA) to obtain an adaptive multi-objective evaluation function M, specifically:
[0158] Initialize the target weight vector W = {w1, w2, ..., w n},Σw i =1; Based on historical performance and current system status, define the importance index I(oi) of each target; Construct adaptive weight update rules:
[0159] w i (t+1) = w i (t) +η* (I(o i ) - w i (t)) +ε*N(0,1);
[0160] Where η is the learning rate, ε is the noise term, and N(0, 1) is the standard normal distribution.
[0161] Apply the Softmax function to normalize the updated weights; construct the fuzzy membership function μ i (x), represents the solution x for target o iThe degree of satisfaction; define the adaptive multi-objective evaluation function: M(x) =Σ(w i *μ i (x)). Periodically evaluate and adjust the weight update strategy.
[0162] The grid operation constraints E and the adaptive multi-objective evaluation function M are input into the dynamic constraint processing and adaptive penalty algorithm (DCAPA) to obtain the constrained adaptive multi-objective evaluation function M c , specifically:
[0163] The constraints E are classified into hard constraints EH and soft constraints ES. For hard constraints EH, the feasibility rules are used to quickly screen solutions that meet the hard constraints, build repair operations, and map infeasible solutions to the feasible domain boundary. For soft constraints ES, the violation degree function v(x, e i ), construct an adaptive penalty factor:
[0164] p(e i , t) = p0(e i ) * (1 +γ* t) δ ;
[0165] Where t is the number of iterations, γ and δ are adjustment parameters, p0(e i ) represents the initial penalty factor.
[0166] Constructing a constrained evaluation function: M c (x) = M(x) -Σ(p(e i , t) * v(x, e i )); Dynamically adjust the penalty strategy, track the history of constraint violations, increase the penalty weight for frequently violated constraints, and if the power grid operation constraint is violated more than 10 times in a month, the constraint is considered to be frequently violated. Regularly re-evaluate the constraint classification and switch between hard and soft constraints when necessary.
[0167] Using the constrained adaptive multi-objective evaluation function M c and the sorted candidate action list R C , the Co-evolutionary Multi-Particle Swarm Optimization Algorithm (CEMPSO) is used to obtain the Pareto optimal solution set P, which is:
[0168] Initialize multiple subpopulations {P1, P2, ..., P k}, each focusing on a specific target or search area. In each subpopulation P i , using the improved velocity update rule:
[0169] v(t+1) = wv(t) + c1r1*(pbest - x) + c2r2(gbest - x) + c3r3(sbest -x);
[0170] Where sbest is the optimal solution of the sub-swarm, w is the adaptive inertia weight, c1, c2, c3 represent acceleration coefficients, r1, r2 and r3 represent random numbers in the range [0, 1], pbest represents the historical optimal position of the particle itself, and gbest represents the global optimal position of the entire particle swarm.
[0171] Introduce a co-evolution mechanism, regularly exchange elite individuals between subpopulations, and use a competition-cooperation strategy to evaluate individual fitness. Construct a diversity preservation strategy, use crowding distance to maintain the diversity of the Pareto frontier, introduce small probability mutations, and jump out of the local optimum. Dynamically adjust the number and size of subpopulations, merge similar subpopulations, and split subpopulations with strong exploratory potential. Apply an archiving strategy to maintain the global non-dominated solution set. Regularly reconstruct the search space, focus on potential areas, and output the final Pareto optimal solution set P.
[0172] Input the Pareto optimal solution set P into the fuzzy analytic hierarchy network (FHAN) to obtain the optimal action A opt , specifically:
[0173] Construct a decision hierarchy, including the target layer, the criterion layer and the scheme layer. In each layer, establish a fuzzy comparison matrix A = (a ij )n×n:a ij = (l ij , m ij ,u ij ), represents a fuzzy number, where l ij , m ij ,u ij Respectively represent the lower bound, median and upper bound of the fuzzy number in the fuzzy comparison matrix. Use the improved eigenvector method to calculate the weights, construct the fuzzy consistency matrix R, solve the fuzzy characteristic equation: RΘW=λmaxΘWc, and reduce the fuzzy weights, where Θ represents the multiplication operator of the fuzzy matrix, W represents the weight vector, λ represents the eigenvalue in the fuzzy characteristic equation, max represents the maximum value operation, and Wc represents the normalized weight vector. Introduce a neural network to process the relationship between layers, use the BP network to learn the mapping relationship between layers, and apply dropout to prevent overfitting. Construct an adaptive membership function:
[0174] μ(x) = 1 / (1 + ((xc) / a) 2 b);
[0175] Among them, a, b, and c are dynamic parameters, which are adjusted according to historical data.
[0176] Use Choquet fuzzy integral to aggregate multiple criteria evaluation: E(f) =Σ(f(x(i)) - f(x(i-1)))*μ(A(i)), where f(x(i) represents the evaluation function value corresponding to criterion i;
[0177] Apply sensitivity analysis to evaluate the robustness of the results and output the action with the highest comprehensive score as the optimal action A opt .
[0178] In another specific embodiment, under multiple optimization objectives, the best action is selected from the sorted candidate actions. Based on the optimization objective set O, a weighted sum method is used to construct an evaluation function for each optimization objective, and the evaluation function is combined into a multi-objective optimization problem to form a multi-objective evaluation function M. Based on the power grid operation constraint condition E and the multi-objective evaluation function M, a penalty function method is used to apply the power grid operation constraint condition to the optimization problem to obtain a constrained multi-objective evaluation function M. c Based on the constrained multi-objective evaluation function M c and the sorted candidate action list R C , the improved multi-objective particle swarm optimization (MOPSO) is used to solve the optimal action and obtain the Pareto optimal solution set P. From the Pareto optimal solution set P, the TOPSIS decision method is used to select the final optimal action A opt . By using graph attention networks and sequential graph neural networks, the topological structure and temporal dynamic characteristics of the power grid are fully considered; a multi-objective optimization method is adopted to balance multiple optimization objectives that may conflict with each other; the concept of Pareto optimality is introduced to provide decision makers with a series of non-dominated solutions; the TOPSIS method is used for the final decision, considering the comprehensive performance of multiple objectives.
[0179] It not only takes into account the complexity and dynamics of the power grid, but also provides a flexible decision-making framework that can adjust the weights of optimization objectives according to different operational needs. The optimal action A is finally output opt It will be the comprehensive optimal choice under multiple objectives and constraints.
[0180] This embodiment improves the overall effect of power grid topology optimization through adaptive multi-objective optimization and fuzzy comprehensive decision-making methods. The application of dynamic weight adaptive fusion algorithm realizes dynamic trade-offs of multiple optimization objectives by defining target importance indicators and constructing adaptive weight update rules. Not only does it take into account the relative importance of different objectives, but it can also make adaptive adjustments based on historical performance and current system status, ensuring the flexibility and adaptability of the optimization process. The introduction of dynamic constraint processing and adaptive penalty algorithm effectively handles complex power grid operation constraints by distinguishing hard constraints from soft constraints and constructing adaptive penalty factors. It not only ensures that the optimization results meet key operating requirements, but also improves the feasibility and practicality of optimization through flexible processing of soft constraints. The use of co-evolutionary multi-population particle swarm optimization algorithm improves the efficiency and diversity of search by initializing multiple sub-populations focusing on different objectives or search areas and introducing co-evolution mechanism. Improved speed update rules and adaptive parameter adjustment further enhance the convergence and exploration capabilities of the algorithm. The application of diversity retention strategy and archiving mechanism ensures the quality and distribution of the Pareto optimal solution set obtained. The application of fuzzy hierarchical analysis network has realized the effective processing of complex decision-making problems by constructing a decision hierarchy and using a fuzzy comparison matrix. The introduction of improved eigenvector method and neural network has improved the accuracy of weight calculation and the modeling ability of inter-layer relationships. The use of adaptive membership function and Choquet fuzzy integral has further enhanced the flexibility and comprehensiveness of decision-making.
[0181] This embodiment achieves efficient solution of power grid topology optimization. Not only can it consider multiple optimization objectives and complex constraints at the same time, but it also effectively handles the uncertainty and complexity in the optimization process through co-evolution and fuzzy comprehensive decision-making methods. This method improves the quality and reliability of the optimization results, and also achieves rapid optimization of large-scale power grid systems through parallel computing and adaptive mechanisms, providing strong support for practical applications. In addition, the flexibility and adaptability of the solution enable it to effectively cope with various complex situations in power grid operation and improve the overall performance and reliability of the system.
[0182] This application significantly improves the overall effect of power grid topology optimization through innovative data preprocessing, feature fusion, action evaluation and multi-objective optimization methods. Data preprocessing and action space construction lay a solid foundation for subsequent optimization, and improve data quality and retrieval efficiency through multi-level action space optimization and structured storage. The dynamic dual-tower feature fusion recall model realizes accurate perception of the power grid state and efficient action recall. Through multi-scale time series feature extraction and adaptive dual-tower model, the accuracy of feature representation and the efficiency of recall are improved. The graph attention and time series graph neural network action evaluation method realizes a comprehensive evaluation of candidate actions through spatiotemporal adaptive graph construction and multi-criteria decision network. The adaptive multi-objective optimization and fuzzy comprehensive decision-making method effectively handles multi-objective optimization problems through dynamic weight adjustment and co-evolution algorithm. Through the organic combination of the above steps, not only the quality and reliability of the optimization results are improved, but also the rapid optimization of large-scale power grid systems is realized through parallel computing and adaptive mechanisms, providing strong support for practical applications.
[0183] In another embodiment of the present application, the power grid topology optimization method based on search sorting may also be:
[0184] S1, multi-source heterogeneous data acquisition and fusion preprocessing;
[0185] S11, real-time power grid topology data acquisition and preprocessing;
[0186] S111. Obtain the grid topology data in real time from the SCADA system, including node connection relationships, switch status, and line parameters, to form the original topology data set D topo_raw ;
[0187] S112, apply the improved Kalman filter algorithm to the original topological data set D topo_raw Perform noise elimination and state estimation to obtain the cleaned topological data set D topo_clean ;
[0188] S113, use a new graph compression algorithm to compress the topological dataset D topo_clean Perform lossless compression to generate compressed topological dataset D topo_compressed ;
[0189] S114, compress the topological data set D topo _ compressed Stored in a distributed graph database to form a structured topological dataset D topo _ structured , which is convenient for subsequent rapid retrieval and analysis;
[0190] S12, multi-dimensional load data collection and prediction;
[0191] S121, collect user-level load data from smart meters and load management systems in real time to form a real-time load data set D load _ real ;
[0192] S122. Extract historical load data from the historical database, including power consumption information at different time scales (hours, days, weeks, and months), to form a historical load data set D load _ hist ;
[0193] S123. Integrate external data sources, such as weather forecasts, socioeconomic indicators, and special event information, to form an external factor dataset D external ;
[0194] S124, the data set D load _ real , D load _ hist and D external The input is fed into the innovative multi-scale attention mechanism combined with the deep learning model of wavelet transform to generate short-term, medium-term and long-term load forecast results, forming the load forecast dataset D load _ pred ;
[0195] S125, the data set D load _ real , D load _ hist and D load _ pred Merge and store in the time series database to form a comprehensive load data set D load _ comprehensive ;
[0196] S13, acquisition and processing of dynamic characteristic data of generator sets;
[0197] S131, real-time collection of operating parameters of the generator set, including power output, speed, excitation current, etc., to form the original generator data set D gen _ raw ;
[0198] S132, using a new nonlinear system identification algorithm, with data set D gen _ raw As input, the dynamic model parameters of the generator set are estimated in real time to generate the dynamic model parameter set D gen _ model ;
[0199] S133, apply the improved particle filter algorithm, combined with data set D gen _ raw and dataset D gen_ model , real-time tracking and prediction of the generator set status, generating a generator status prediction data set D gen _ pred ;
[0200] S134, the data set D_ en _ raw , D gen _ model and D gen _ pred Merge and store in a high-performance key-value storage system to form a comprehensive generator dataset D gen _ comprehensive ;
[0201] S14, power grid state estimation and data fusion;
[0202] S141, collect real-time measurement data such as voltage, current, phase angle, etc. from various sensors and measurement devices to form an original measurement data set D meas _ raw ;
[0203] S142, using innovative distributed state estimation algorithms, with dataset D meas _ raw and dataset D topo _ structured As input, combined with edge computing technology, fast and accurate grid state estimation is achieved to generate a state estimation result dataset D state _ est ;
[0204] S143, using a new multi-source data fusion algorithm, the data set D topo _ structured , D load _ comprehensive , D gen _ comprehensive and D state _ est Perform comprehensive processing to generate a preliminary fusion data set D fusion _ prelim ;
[0205] S144, Apply the deep learning-based anomaly detection algorithm to the dataset D fusion _ prelim As input, identify and label potential data anomalies and measurement errors, and generate a fused dataset D with anomaly labels fusion _ anomaly ;
[0206] S145, the data set D fusion _ anomalyStored in a distributed spatiotemporal database to form the final grid state fusion data set D fusion _ final , supporting efficient spatiotemporal query and analysis;
[0207] S15, feature engineering and data enhancement;
[0208] S151. From data set D fusion _ final Extract the original features, including node features, edge features and global features, to form the original feature set D feature _ raw ;
[0209] S152, using a new graph kernel method, with dataset D feature _ raw and D topo _ structured As input, high-order features that reflect the topological characteristics of the power grid are generated to form a high-order feature set D feature _ high ;
[0210] S153, using the improved variational autoencoder, with data set D feature _ raw and D feature _ high As input, learn the implicit representation of the power grid state and generate the implicit feature set D feature _ latent ;
[0211] S154, apply innovative conditional generative adversarial network (CGAN) variants to dataset D fusion _ final As conditional input, a variety of power grid operation scenarios are generated to form an enhanced data set D augmented ;
[0212] S155, using the feature selection algorithm based on mutual information, from the data set D feature _ raw , D feature _ high and D feature _ latent The most representative and discriminative feature subset is selected to form the final feature set D feature _ final ;
[0213] S156, the data set D feature _ final and D augmented Merge and store in the feature storage system to form a comprehensive feature data set D feature _ comprehensive, supporting subsequent model training and decision-making processes;
[0214] S16. Data quality assessment and metadata management;
[0215] S161. Build and implement an innovative data quality scoring system based on dataset D fusion _ final and D feature _ comprehensive As input, the data quality score set D is generated by comprehensively considering multiple dimensions such as completeness, consistency, accuracy and timeliness. quality _ score ;
[0216] S162. Apply the data traceability mechanism based on hash chain to record the data from the original data to the data set D fusion _ final and D feature _ comprehensive The whole process ensures the traceability and auditability of data and generates data traceability information set D provenance ;
[0217] S163, using a new metadata automatic extraction and annotation algorithm, based on dataset D fusion _ final and D feature _ comprehensive As input, generate detailed data description information to form metadata set D metadata ;
[0218] S164, Apply knowledge graph technology to dataset D metadata Based on this, we construct a knowledge graph of power grid data, realize the semantic association between data, and generate a knowledge graph dataset D knowledge _ graph ;
[0219] S165, the data set D quality _ score , D provenance , D metadata and D knowledge _ graph Merge and store in a dedicated metadata management system to form a comprehensive metadata set D metadata _ comprehensive , support data governance and management decisions;
[0220] S17, data preprocessing pipeline construction and optimization;
[0221] S171. Build a configurable data preprocessing pipeline based on data set D fusion _ final , D feature _ comprehensiveand D metadata _ comprehensive As input, it includes modules such as data cleaning, standardization, missing value processing and outlier detection;
[0222] S172, apply a new adaptive data cleaning algorithm, according to the data set D metadata _ comprehensive The data characteristics in the data automatically select the most suitable cleaning strategy, clean the input data, and generate the cleaned data set D clean ;
[0223] S173. Use the improved multi-task learning framework to process dataset D simultaneously clean The data interpolation and anomaly detection tasks in the above example generate the dataset D after interpolation and anomaly detection. imputed _ detected ;
[0224] S174, apply the hyperparameter tuning algorithm based on Bayesian optimization to automatically adjust the parameters of the preprocessing pipeline and optimize the data set D imputed _ detected quality;
[0225] S175, realize parallelization and distributed processing of preprocessing pipeline, improve data set D imputed _ detected processing speed and efficiency;
[0226] S176, the final processed high-quality data set D final Stored in a distributed file system as the basic data source for subsequent analysis and modeling.
[0227] S2, dynamic topological space construction and optimization;
[0228] S21, initial topological space generation;
[0229] S211, obtaining the final processed high-quality data set D from step S1 final , extract the current grid topology T current ;
[0230] S212, based on the grid topology T current , use the improved graph traversal algorithm to enumerate all possible switch operations and generate a candidate topology set T candidates ;
[0231] S213, applying a fast constraint checking algorithm to the candidate topology set T candidates As input, the topological structure that meets the basic physical constraints is screened out to form the initial feasible topological space T feasible ;
[0232] S214, use the graph isomorphism algorithm to initialize the feasible topological space T feasible The topological structure in is deduplicated to obtain a simplified feasible topological space T unique ;
[0233] S22, dynamic update of topological space;
[0234] S221, construct an adaptive sampling strategy based on the feasible topological space T unique and historical topological effect data H effects As input, generate sampling probability distribution P sample ;
[0235] S222, using the improved Monte Carlo tree search (MCTS) method, with the feasible topological space T unique and the sampling probability distribution P sample As input, explore the topological space and generate the extended topological set T extended ;
[0236] S223, apply marginal benefit analysis algorithm to evaluate the extended topology set T extended The potential benefits of each topological structure in the process are calculated to generate the benefit evaluation result R marginal ;
[0237] S224, based on the benefit evaluation results R marginal , dynamically adjust the size of the topological space, prune low-yield topologies, and form an optimized topological space T optimized ;
[0238] S23, topology evaluation index calculation;
[0239] S231, using fast power flow calculation algorithm, with topological space T optimized and D final The load data in is used as input to calculate the power loss of each topology and generate the loss index set I loss ;
[0240] S232, Apply the improved voltage stability margin assessment method to calculate the topological space T optimized The stability index of each topology in forms a stability index set I stability ;
[0241] S233, combine the equipment life model and maintenance cost data to calculate the topological space T optimized The economic index of each topology in is obtained, and the economic index set I is obtained. economy ;
[0242] S234, Integrated indicator set I loss ,I stability and I economy, generate comprehensive evaluation index set I comprehensive ;
[0243] S24, topological space representation;
[0244] S241, construct innovative graph embedding algorithms based on topological space T optimized As input, each topological structure is converted into a low-dimensional vector representation to generate a topological embedding set E topology ;
[0245] S242. Use the improved attention mechanism to analyze the topological space T optimized The importance of the nodes and edges in the , generating the importance weight set W importance ;
[0246] S243, combined with topological embedding set E topology and the importance weight set W importance , construct multi-scale topological features and generate feature set F multiscale ;
[0247] S244, apply dimensionality reduction techniques (such as t-SNE) to transform the feature set F multiscale Map to the visualization space and generate a visualization dataset V topology ;
[0248] S25, Topological space clustering and representative selection;
[0249] S251, using the spectral clustering algorithm, the feature set is F multiscale As input, for the topological space T optimized Perform clustering and obtain the clustering result C topology ;
[0250] S252, construct a representative selection algorithm based on centrality, select the most representative topological structure from each cluster, and form a representative topological set T representative ;
[0251] S253, apply diversity preservation algorithm to ensure the representative topology set T representative Cover different aspects of the topological space and generate a diverse representative topological set T diverse ;
[0252] S26, topology conversion operation sequence generation;
[0253] S261, construct the minimum edit distance algorithm and calculate the topological set T current to T diverse The conversion path of each topology in generates the initial conversion sequence set S transform _ initial ;
[0254] S262, apply the path optimization algorithm, consider the switching operation cost and system stability, and optimize the initial conversion sequence set S transform _ initial , and get the optimized transformation sequence set S transform _ optimized ;
[0255] S263, use sequence compression algorithm to simplify the conversion sequence set S transform _ optimized The operation steps in generate the final transformation operation sequence set S transform _ final ;
[0256] S27, dynamic topological spatial index construction;
[0257] S271. Construct an efficient graph index structure based on the representative topological set T diverse As input, construct a topological space index I topology ;
[0258] S272, using incremental update algorithm to ensure the topological space index I topology Able to dynamically adapt to T diverse changes;
[0259] S273, implement topological space index I topology The fast query mechanism supports efficient topological retrieval in subsequent steps;
[0260] S28, Topological spatial data packaging and storage;
[0261] S281, the representative topological set T diverse , feature set F multiscale , Comprehensive evaluation index set I comprehensive , Visualization Dataset V topology , conversion operation sequence set S transform _ final and topological space index I topology Packaging to form a comprehensive topological space data package D topology _ package ;
[0262] S282, using data serialization technology, the data packet D topology _ package Convert to a format that is easy to store and transmit;
[0263] S283, storing the serialized data in a distributed storage system to form a persistent topological spatial dataset D topology _ persistent , providing a basis for subsequent optimization and decision-making processes.
[0264] S3, construction and solution of multi-objective optimization model;
[0265] S31, objective function construction and formulation construction;
[0266] S311, obtain the persistent topological spatial dataset D from step S2 topology _ persistent , extract comprehensive evaluation index set I comprehensive ;
[0267] S312, based on comprehensive evaluation index set I comprehensive Loss index I loss , construct the power loss minimization objective function F loss ;
[0268] S313. Use comprehensive evaluation index set I comprehensive Stability index I stability , construct the system stability maximization objective function F stability ;
[0269] S314, according to the comprehensive evaluation index set I comprehensive Economic indicators in I economy , formulate the economic optimization objective function F economy ;
[0270] S315, introduce the environmental impact assessment model, consider factors such as carbon emissions, and construct the environmentally friendly objective function F environment ;
[0271] S316, integration function F loss 、F stability 、F economy and F environment , forming the objective function set F of the multi-objective optimization problem objectives ;
[0272] S32, Constraint Modeling and Integration;
[0273] S321. From data set D topology _ persistent Extract the topological structure information T diverse , combined with the power system physical model, establish the node power balance constraint C power _ balance ;
[0274] S322, based on data packet D final The equipment parameter information in the line capacity constraint C is constructed line _ capacity ;
[0275] S323, using voltage stability theory, set voltage and frequency stability constraints C voltage _ frequency ;
[0276] S324. Considering the equipment operation life and reliability requirements, add equipment operation times limit constraint C operation _ limit ;
[0277] S325, Integration Constraint C power _ balance , C line _ capacity , C voltage _ frequency and C operation _ limit , forming a constraint set C constraints ;
[0278] S33, Problem Formulation and Preprocessing;
[0279] S331, the function F objectives and constraint C constraints Combine and formulate the complete multi-objective optimization problem formulation P multi _ objective ;
[0280] S332, apply the objective function normalization technology to ensure that the function F objectives The scale consistency of each objective function in generates a normalized objective function set F normalized ;
[0281] S333, use constraint preprocessing technology to simplify constraint C constraints The redundant constraints in the result are reduced to obtain the set of constraints C. simplified ;
[0282] S334, combined function F normalized and constraint C simplified , forming the optimization problem formulationP after preprocessing preprocessed ;
[0283] S34, Construction and implementation of multi-objective optimization algorithm;
[0284] S341. Construct an improved multi-objective evolutionary algorithm (MOEA) framework, including individual encoding schemes, crossover and mutation operators, to adapt to the characteristics of topology optimization problems.
[0285] S342, implement a fast non-dominated sorting algorithm based on dominance relations to evaluate the Pareto advantage of solutions;
[0286] S343, introduce adaptive weight adjustment mechanism, dynamic balance function F normalized Multiple targets in the equation generate a weight vector set W adaptive ;
[0287] S344. Construct a solution improvement strategy based on local search to improve the local optimization ability of the algorithm;
[0288] S345, implement a parallel computing framework to support efficient solution of large-scale optimization problems;
[0289] S35, initial solution generation and population initialization;
[0290] S351, from data set D topology _ persistent Extract the representative topological set T representative , as the candidate set of the initial solution;
[0291] S352, apply the greedy algorithm, the basis function is F normalized Quickly generate a set of initial feasible solutions S initial ;
[0292] S353, using Latin hypercube sampling method, uniformly sample in the solution space to generate a diversified initial solution set S diverse ;
[0293] S354, merge solution set S initial and S diverse , forming the initial population P initial ;
[0294] S36, Iterative Optimization and Convergence Control;
[0295] S361, using the constructed MOEA framework, with population P initial As input, start the iterative optimization process;
[0296] S362, in each iteration, crossover and mutation operations are applied to generate new candidate solutions to form the offspring population P offspring ;
[0297] S363, combined with the parent population P parent and the offspring population P offspring , using fast non-dominated sorting and crowding calculation, select the new generation population P new ;
[0298] S364, applying the adaptive weight adjustment mechanism to the weight vector set W adaptive , dynamically adjust the weight of the objective function;
[0299] S365, regularly execute local search strategy to improve the new generation population P newThe non-dominated solutions in ;
[0300] S366, checking the convergence condition and the number of iterations, and deciding whether to continue the iteration or terminate the optimization process;
[0301] S37, Pareto frontier solution set generation and evaluation;
[0302] S371, from the final population P final Extract non-dominated solutions from the Pareto frontier solution set S pareto ;
[0303] S372, using the improved hypervolume indicator, evaluate the frontier solution set S pareto The quality and diversity of the generated evaluation index I hypervolume ;
[0304] S373, apply clustering algorithm to the frontier solution set S pareto Group and select representative solutions to form a simplified Pareto solution set S pareto _ representative ;
[0305] S374, perform sensitivity analysis of the solution and evaluate the solution set S pareto _ representative The robustness of each solution in generates a robustness index set I robustness ;
[0306] S38, solution decision support and visualization;
[0307] S381. Construct an interactive decision support system that allows decision makers to choose from a set of solutions based on their preferences. pareto _ representative Choose from;
[0308] S382, implement the solution sorting method based on fuzzy reasoning to provide decision makers with a recommended solution set S recommended ;
[0309] S383. Use high-dimensional data visualization techniques, such as parallel coordinates, to intuitively display the solution set S pareto _ representative The distribution characteristics of
[0310] S384, generate a decision report, including optimization results, performance indicators and sensitivity analysis, to form a comprehensive decision support package D decision _ support ;
[0311] S39, optimization result storage and knowledge accumulation;
[0312] S391, the solution set S pareto 、Solution set Spareto _ representative 、Indicator I hypervolume 、Indicator I robustness and support package D decision _ support Packaging to form optimization result data package D optimization _ results ;
[0313] S392, build an incremental learning mechanism, from data packet D optimization _ results Extract new knowledge and update the optimization strategy knowledge base K optimization ;
[0314] S393, data packet D optimization _ results Stored in a distributed database to form a historical optimization result set H optimization _ history ;
[0315] S394, update the historical topology effect data H in step S2 effects , providing feedback for the subsequent dynamic update of the topological space.
[0316] S4, real-time decision-making and execution control;
[0317] S41, real-time status monitoring and data acquisition;
[0318] S411. Obtain current power grid operation status data from the SCADA system in real time current _ state , including topology, load distribution, power generation, etc.;
[0319] S412, using the data preprocessing pipeline in step S1, the previous power grid operation status data D current _ state Perform rapid cleaning and standardization to obtain processed real-time status data D current _ processed ;
[0320] S413, from the historical optimization result set H of step S3 optimization _ history Extract the most recent optimization results as reference data D recent _ optimization ;
[0321] S42, rapid decision-making model construction;
[0322] S421. Based on reference data D recent _ optimization , use ensemble learning methods (such as random forests, gradient boosting trees) to build a fast decision model Mfast _ decision ;
[0323] S422. Use real-time status data D current _ processed As input, through the fast decision model M fast _ decision Generate preliminary decision recommendations initial _ decision ;
[0324] S423, apply the confidence assessment algorithm to calculate the preliminary decision recommendation D initial _ decision Credibility Index I confidence ;
[0325] S43, Application of deep reinforcement learning model;
[0326] S431, from the knowledge base K optimization Load the pre-trained deep reinforcement learning model M in DRL ;
[0327] S432, use real-time status data D current _ processed As state input, through the deep reinforcement learning model M DRL Generate action suggestions A DRL ;
[0328] S433, Calculation Action Suggestion A DRL The Q value is used as the action quality indicator I Q_value ;
[0329] S44, Decision Fusion and Optimization;
[0330] S441. Construct a decision fusion algorithm based on fuzzy logic and comprehensively consider preliminary decision suggestions D initial _ decision 、Action suggestion A DRL , Credibility Index I confidence and action quality index I Q _ value ;
[0331] S442, generate fused decision suggestions D fused _ decision ;
[0332] S443, apply fast verification algorithm to ensure decision-making recommendations D fused _ decision Satisfy key constraints and form feasible decisions D feasible _ decision ;
[0333] S45. Safety assessment and emergency response;
[0334] S451, using digital twin technology, based on real-time status data current _ processed and feasible decision D feasible _ decision , quickly simulate the grid status S after decision execution simulated ;
[0335] S452. Apply security assessment model to analyze power grid status S simulated Potential risks, generate risk assessment report R risk ;
[0336] S453, if the risk assessment report R risk There are high-risk items in the system, triggering the emergency decision-making process and generating emergency plans D emergency ;
[0337] S46, generating and issuing execution instructions;
[0338] S461, based on feasible decision D feasible _ decision Or emergency plan D emergency (if triggered), generates a detailed set of execution instructions I execution ;
[0339] S462, using natural language processing technology, will execute instruction set I execution Convert to human-readable instructions operation ;
[0340] S463, through the secure communication channel, execute instruction set I execution Issued to relevant control equipment and operators;
[0341] S47, execution monitoring and feedback collection;
[0342] S471, real-time monitoring of instruction execution process, collection of execution status data D execution _ status ;
[0343] S472, compare the actual power grid state after execution with the power grid state S simulated , generate execution effect evaluation report R execution ;
[0344] S473. Collect feedback from operators and generate manual feedback data D human _ feedback ;
[0345] S48, decision-making effect evaluation and model updating;
[0346] S481, Comprehensive analysis and implementation effect evaluation report R execution and manual feedback data D human _ feedback , generate decision effect evaluation report R decision _ effect ;
[0347] S482, based on decision-making effect evaluation report R decision _ effect , update the fast decision model M fast _ decision and deep reinforcement learning model M DRL Online learning parameters of
[0348] S483, adding new decision experience to the experience playback buffer D experience _ buffer middle;
[0349] S49, knowledge accumulation and experience inheritance;
[0350] S491, regularly replay the experience buffer D experience _ buffer Extract valuable experience from the knowledge base K and update it optimization ;
[0351] S492. Generate a summary report of the decision-making process R decision _ summary , stored in the decision history database H decision _ history ;
[0352] S493, update the historical optimization result set H in step S3 optimization _ history , providing the latest feedback for subsequent optimization.
[0353] S5. System maintenance and continuous optimization;
[0354] S51, performance monitoring and diagnosis;
[0355] S511. Continuously collect the operating performance data of each module of the system performance , including response time, resource utilization, etc.;
[0356] S512. Use anomaly detection algorithm to analyze operation performance data D performance , identify potential performance bottlenecks and anomalies, and generate diagnostic reports R diagnosis ;
[0357] S513, based on the diagnosis report R diagnosis , generate a list of system optimization suggestions L optimization _suggestions ;
[0358] S52, data quality management;
[0359] S521. Regularly evaluate the quality of various types of data in the system and generate data quality reports. data _ quality ;
[0360] S522, build data quality improvement strategy, and report data quality data _ quality Problems in the development of improvement plans data _ improvement ;
[0361] S523. Implement improvement plan P data _ improvement , and track the improvement effects and update the data processing process;
[0362] S53, model performance evaluation and update;
[0363] S531, Regular evaluation of rapid decision-making model M fast _ decision and deep reinforcement learning model M DRL Performance, generate model evaluation report R model _ evaluation ;
[0364] S532, Model-based evaluation report R model _ evaluation and the latest historical optimization result set H optimization _ history , construct the model update strategy S model _ update ;
[0365] S533. Implementing model update strategy in offline environment model _ update , generate the updated model fast decision model M fast _ decision _ new and deep reinforcement learning model M DRL _ new ;
[0366] S534. Conduct A / B testing to compare the performance of the new and old models and decide whether to deploy the updated version;
[0367] S54, Algorithm optimization and innovation;
[0368] S541, analysis and decision history database H decision _ history, identify the limitations of existing algorithms and room for improvement;
[0369] S542. Combine the latest research progress to build improved algorithms and methods A new _ algorithms ;
[0370] S543, Test Method A in a Simulated Environment new _ algorithms , evaluate its performance and feasibility;
[0371] S544. Develop an integration plan for new algorithms with significant performance improvements algorithm _ integration ;
[0372] S55, knowledge base update and maintenance;
[0373] S551. Regular review of knowledge base K optimization , remove outdated or irrelevant knowledge;
[0374] S552. Integrate newly acquired knowledge and experience and update the knowledge base K optimization structure and content;
[0375] S553. Implement knowledge graph technology to optimize the knowledge base K optimization organization and retrieval efficiency;
[0376] S56, system architecture optimization;
[0377] S561. Evaluate the scalability and availability of the current system architecture and generate an architecture evaluation report R architecture ;
[0378] S562, based on the architecture assessment report R architecture , build system architecture optimization solution P architecture _ optimization ;
[0379] S563. Verify the optimization solution in the test environment P architecture _ optimization effectiveness;
[0380] S564. Develop and implement architecture upgrade plans to ensure a smooth transition of the system;
[0381] S57, safety and reliability improvement;
[0382] S571. Conduct regular security audits and generate security assessment reports security ;
[0383] S572, based on the full assessment report R security , formulate a security reinforcement planPsecurity _ enhancement ;
[0384] S573, Implement reinforcement plan P security _ enhancement , and conduct vulnerability scanning and penetration testing;
[0385] S574. Optimize the system's fault recovery mechanism to improve system reliability and resilience;
[0386] S58, user interface and interaction optimization;
[0387] S581. Collect user feedback D user _ feedback , analyze user interaction patterns;
[0388] S582. Based on user feedback D user _ feedback , build a user interface improvement plan P UI _ improvement ;
[0389] S583. Develop and test new user interface prototypes and conduct usability testing;
[0390] S584, gradually roll out interface updates and collect user adaptability data;
[0391] S59, Continuous learning and knowledge updating;
[0392] S591. Track the latest research progress and technological development trends in the field;
[0393] S592. Organize regular internal training and knowledge sharing meetings;
[0394] S593. Establish partnerships with academia and industry to promote knowledge exchange;
[0395] S594. Integrate new knowledge and insights into every aspect of the system to ensure the advancement and adaptability of the system.
[0396] The power grid topology optimization system based on search and sorting includes:
[0397] at least one processor; and,
[0398] a memory communicatively connected to at least one of the processors; wherein,
[0399] 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 topology optimization method based on search sorting described in any one of the above embodiments.
[0400] This application provides new ideas and methods for solving these problems through innovative data preprocessing, dynamic feature fusion, multi-scale graph neural network and adaptive multi-objective optimization technologies, and is expected to achieve breakthrough progress in improving optimization efficiency, enhancing adaptability and improving decision-making quality. The multi-scale time series feature extraction algorithm decomposes the input time series into multiple time scales and uses an adaptive window sliding algorithm to capture features at different scales. This can effectively capture the dynamic changes in the state of the power grid. The spatiotemporal adaptive graph construction algorithm achieves accurate modeling of the changes in the topology of the power grid over time by dynamically adjusting the time window and edge weights. The combination of these two algorithms enables the system to adapt to changes in the state of the power grid in real time, improving the adaptability and accuracy of the optimization results.
[0401] The quantification of target trade-offs in multi-target optimization problems is achieved through a dynamic weight adaptive fusion algorithm. Dynamic trade-offs of multiple optimization targets are achieved by defining target importance indicators and constructing adaptive weight update rules. The system can automatically adjust the weights of each target based on historical performance and current system status, ensuring the flexibility and adaptability of the optimization process. In addition, the application of fuzzy hierarchical analysis networks further improves the accuracy and reliability of multi-target decision-making by constructing a decision hierarchy and using a fuzzy comparison matrix.
[0402] The adaptive dual-tower model and multi-scale dynamic graph attention network achieve efficient screening and evaluation of massive possible actions in large-scale power grid systems. The adaptive dual-tower model achieves efficient indexing and retrieval of large-scale action spaces through dynamic structure multi-layer perceptrons and hierarchical local sensitive hashing. The multi-scale dynamic graph attention network can quickly capture complex spatiotemporal dependencies in the power grid through multiple attention heads and time decay factors, improving the efficiency and accuracy of action evaluation. The combination of these methods enhances the system's ability to handle massive actions and improves optimization efficiency.
[0403] The dynamic constraint processing and adaptive penalty algorithm realize the processing of complex and changeable constraint conditions. By distinguishing hard constraints from soft constraints and constructing adaptive penalty factors, flexible processing of complex constraint conditions is achieved. For hard constraints, the algorithm uses feasibility rules to quickly screen solutions that meet the conditions; for soft constraints, by dynamically adjusting the penalty weights, various operating constraints are met while ensuring optimization efficiency. This method not only ensures the feasibility of the optimization results, but also improves the efficiency of the optimization process.
[0404] The coordination of topology optimization decisions at different time scales is achieved through a multi-scale temporal feature extraction algorithm and an adaptive spatiotemporal graph convolutional network. The multi-scale temporal feature extraction algorithm can capture both short-term and long-term temporal features, while the adaptive spatiotemporal graph convolutional network achieves effective fusion of features at different time scales through a multi-layer network structure, including spatial convolutional layers, temporal convolutional layers, and combination layers. This approach enables the system to consider both short-term and long-term goals at the same time, achieving coordination and consistency of decisions at different time scales.
[0405] This application effectively solves the key technical problems in power grid topology optimization through innovative algorithm construction and the organic combination of multiple methods. It not only improves the optimization efficiency, but also enhances the adaptability and decision-making quality of the system, providing strong support for the practical application and promotion of power grid topology optimization technology.
[0406] 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 topology optimization method based on search sorting, characterized in that: The steps include: S1. Acquire power grid topology data, and construct an action space based on the power grid topology data; wherein the power grid topology data includes a node set, an edge set, and an attribute set; S2, obtaining real-time grid status data, building an adaptive dual-tower model based on the real-time grid status data and action space, and calculating a candidate action set; S3. Based on the candidate action set, a graph attention network and a temporal graph neural network are constructed to obtain a sorted list of candidate actions. S4, based on the sorted candidate action list, perform multi-objective optimization and decision making to obtain the optimal action; Step S3 is further as follows: S31, acquiring historical power grid status data, and constructing a dynamic space-time graph based on the historical power grid status data and power grid topology data; S32. Based on the dynamic spatiotemporal graph and state embedding vector, a multi-scale dynamic graph attention network is constructed to obtain attention-weighted node representation; S33. Based on the attention-weighted node representation and candidate action set, an adaptive spatiotemporal graph convolutional network is constructed to obtain the spatiotemporal representation of action-state pairs. S34, based on the spatiotemporal representation of action-state pairs, construct a hierarchical multi-criteria decision network to generate the final scores of candidate actions and a ranked list of candidate actions; Step S32 is further as follows: S321. Construct a multi-scale dynamic graph attention network based on dynamic spatiotemporal graph and state embedding vector; S322. Based on the state embedding vector, calculate the query, key, and value matrices of each attention head in the multi-scale dynamic graph attention network; S323, based on the query, key, and value matrix, calculate an attention score, introduce a time decay factor to modify the attention score, and obtain a modified attention score; S324, based on the modified attention score and value matrix, calculate the output of each attention head, use the gating mechanism to fuse the output of each attention head, and obtain a fusion result; S325, based on the fusion result and the state embedding vector, using residual connection and layer normalization, calculate the attention-weighted node representation; Step S4 is further as follows: S41, based on a predefined set of optimization objectives, a dynamic weight adaptive fusion algorithm is used to obtain an adaptive multi-objective evaluation function; wherein the set of optimization objectives includes load balancing, network stability and energy efficiency; S42, based on the preset power grid operation constraints and the adaptive multi-objective evaluation function, a dynamic constraint processing and an adaptive penalty algorithm are used to obtain an adaptive multi-objective evaluation function with constraints; S43, based on the constrained adaptive multi-objective evaluation function and the ranked candidate action list, a co-evolutionary multi-population particle swarm optimization algorithm is used to obtain the Pareto optimal solution set; S44. Based on the Pareto optimal solution set, a fuzzy hierarchical analysis network is constructed to obtain the action with the highest comprehensive score as the optimal action; Step S41 is further as follows: S411, obtaining an optimization target set, and initializing a target weight vector of the optimization target set; S412, obtaining historical performance data and current system status data, and constructing an importance index for each target based on the historical performance data and the current system status data; S413, based on the importance index, updating the target weight vector to obtain an updated weight, and normalizing the updated weight using a Softmax function to obtain a normalized weight; S414, constructing a fuzzy membership function, and obtaining an adaptive multi-objective evaluation function based on the fuzzy membership function and the normalized weights; Step S42 is further as follows: S421, obtaining preset power grid operation constraint conditions, and classifying the preset power grid operation constraint conditions into hard constraints and soft constraints; S422, based on the hard constraints, using feasibility rules to quickly screen the adaptive multi-objective evaluation function to obtain an adaptive multi-objective evaluation function that satisfies the hard constraints; S423. Based on soft constraints, construct a default degree function and an adaptive penalty factor; S424, constructing a constrained adaptive multi-objective evaluation function based on the adaptive multi-objective evaluation function that satisfies the hard constraints, the breach degree function and the adaptive penalty factor; S425, collecting historical data on violations of preset power grid operation constraints, adding penalty weights to frequently violated preset power grid operation constraints, and obtaining preset power grid operation constraints after weight adjustment; Based on the preset grid operation constraints after adjusting the weights, the classification of hard constraints and soft constraints is dynamically adjusted.
2. The power grid topology optimization method based on search sorting according to claim 1 is characterized in that: Step S1 is further as follows: S11, acquiring power grid topology data and action data affecting the power grid topology; the power grid topology data includes a node set, an edge set and an attribute set; the action data includes an action type, an affected device ID set and an execution parameter set; S12, using a graph traversal algorithm to process the power grid topology data, and generating a basic action set based on the action data and the processed power grid topology data; S13, based on the basic action set, using a combinatorial optimization algorithm to form an action combination set; based on the action combination set, using a heuristic algorithm to obtain an optimized action space; S14, based on the optimized action space and the predefined constraint condition set, sequentially using the constraint propagation algorithm and the decision tree algorithm to obtain the classified action space; based on the classified action space, using the heuristic rule engine to obtain the final optimized action space; S15. The finally optimized action space is structured and stored to form a structured data set.
3. The power grid topology optimization method based on search sorting according to claim 2 is characterized in that: Step S2 is further as follows: S21, acquiring real-time power grid status data, and using a multi-scale time series feature extraction algorithm based on the real-time power grid status data to obtain a dynamic feature vector; S22, based on the power grid topology data, using the adaptive spectral clustering enhancement algorithm to obtain a static feature vector; S23, based on the dynamic feature vector and the static feature vector, adopting an adaptive weighted fusion algorithm to obtain a fused feature vector; S24, constructing an adaptive dual-tower model, inputting the fused feature vector into the state tower in the dual-tower model to obtain a state embedding vector; inputting the finally optimized action space into the action tower in the dual-tower model to obtain an action embedding vector set; S25. Based on the state embedding vector and the action embedding vector set, an adaptive similarity calculation and a dynamic threshold selection algorithm are used to obtain a candidate action set.
4. The power grid topology optimization method based on search sorting according to claim 3 is characterized in that: Step S21 is further as follows: S211, acquiring real-time power grid status data, and extracting time series from the real-time power grid status data; S212, decomposing the time series into predetermined time scales, using an adaptive window sliding algorithm for each time scale, and calculating the size of each window; S213, based on the real-time power grid status data, calculating statistical features in each window, and using wavelet transform to extract frequency domain features in the statistical features; The statistical characteristics include mean, variance and kurtosis; S214, constructing a feature importance evaluation index, and updating the weight of the frequency domain feature based on the feature importance evaluation index; S215: concatenate the frequency domain features after the weights are updated to form a final dynamic feature vector.
5. A power grid topology optimization system based on search sorting, characterized in that: 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 topology optimization method based on search sorting as described in any one of claims 1 to 4.
Citation Information
Cited By
Power network topology optimization method based on time series data analysis
CN119787351A
Power Grid Topology Optimization Method Based on Time-Series Data Analysis
CN119787351B
