Power grid emergency resource scheduling optimization method based on source grid load storage multi-dimensional data fusion

Abstract

The invention relates to the technical field of emergency resource scheduling, in particular to a power grid emergency resource scheduling optimization method based on source grid load storage multi-dimensional data fusion. The method comprises the following steps: deploying edge computing nodes at key nodes of a power grid, collecting real-time power grid operation data, meteorological data and geographic information data, carrying out spatial-temporal feature fusion by using a graph neural network, and optimizing the spatial-temporal feature fusion process of the graph neural network by using a median aggregation method; based on the fused spatio-temporal characteristics, load prediction is carried out by adopting a long-short-term memory network model; and constructing an emergency resource scheduling optimization model through a dynamic planning method based on a prediction result, and generating an optimal resource scheduling path. According to the design of the invention, power grid operation data, meteorological data, geographic information and equipment state data are integrated, spatial-temporal feature fusion is carried out by using the graph neural network, and the method can more accurately capture the complex structure and dynamic change of a power grid system.

Description

Technical Field

[0001] The present invention relates to the technical field of emergency resource scheduling, and more specifically, to an optimization method for power grid emergency resource scheduling based on multi-dimensional data fusion of source-network-load-storage. Background Art

[0002] The power grid system consists of different types of nodes (such as substations, distribution terminals, sensor networks, etc.) and connection relationships. These nodes and edges have different attributes and functions. Directly performing graph convolution operations may ignore the essential differences between nodes and fail to accurately capture the complex structure and dynamic changes of the power grid. During the operation of the power grid, a large amount of real-time data is generated, including power grid operation data such as current, voltage, and power, meteorological data such as wind speed, temperature, and precipitation, and geographical information data such as the topological structure of the power grid and the geographical location of equipment. How to effectively integrate these multi-source heterogeneous data and extract useful information for decision support is a challenge. When a fault or emergency occurs in the power grid, quickly locating the fault point, evaluating the impact range, and formulating an effective resource scheduling plan are crucial. However, existing methods often lack sufficient flexibility and adaptability. When a power outage fault occurs, it leads to an extended power outage time and affects power supply reliability. In addition, with the transformation of the energy structure, the coordinated operation of source-network-load-storage has become an important direction for the development of the power grid. The volatility of the source side (such as new energy power generation), the diversity of the load side (user load), and the dynamic regulation characteristics of energy storage have brought new challenges to power grid emergency resource scheduling. Existing methods mostly focus on the optimization of a single link and lack consideration of the dynamic coupling of multiple links of source-network-load-storage, making it difficult to achieve efficient collaborative scheduling of global resources. Therefore, an optimization method for power grid emergency resource scheduling based on multi-dimensional data fusion of source-network-load-storage is provided. Summary of the Invention

[0003] The purpose of the present invention is to provide an optimization method for power grid emergency resource scheduling based on multi-dimensional data fusion of source-network-load-storage, so as to solve the problems of dynamic changes and heterogeneity of complex power grid systems, the effective integration of multi-source heterogeneous data, and the current situation of slow emergency response speed and low resource scheduling efficiency proposed in the above background art.

[0004] To achieve the above object, the present invention aims to provide an optimization method for power grid emergency resource scheduling based on multi-dimensional data fusion of source-network-load-storage, including the following steps: S1. Deploy edge computing nodes at key nodes of the power grid to collect real-time power grid operation data, meteorological data, and geographical information data; S2. Integrate power grid operation data, meteorological data, geographical information, and equipment status data, perform spatio-temporal feature fusion using a graph neural network, and optimize the process of spatio-temporal feature fusion of the graph neural network by combining node type embedding vectors with real-time state factors and adopting the median aggregation method; S3. Based on the fused spatio-temporal features, use a long short-term memory network model for load forecasting; S4. Based on the prediction results, construct an emergency resource scheduling optimization model through dynamic programming method to generate the optimal resource scheduling path.

[0005] As a further improvement of this technical solution, in S2, integrate power grid operation data, meteorological data, geographical information, and equipment status data, and use a graph neural network for spatio-temporal feature fusion, including the following steps: S2.1. Regard the power grid as a graph ; Among them, represents the nodes in the power grid, represents the connection relationship between nodes; S2.2. For each node merge into a comprehensive feature vector : ; Among them, represents the power grid operation data; represents the meteorological data; represents the geographical information; represents the equipment status data; S2.3. Initialize the feature of each node , integrate all features from different sources into the initial feature , as the input of the graph neural network; S2.4. Perform multi-layer graph convolution operations to form the feature of each node, including the spatial domain feature, meteorological feature, and geographical information feature of the node; S2.5. For the power grid operation data, use the sliding window technique to extract the time series features in the power grid operation data; S2.6. Combine the time series features obtained from time series analysis with the spatial domain features obtained through the graph neural network to form spatio-temporal features.

[0006] As a further improvement of this technical solution, in S2.4, perform multi-layer graph convolution operations, including the following steps: S2.41. Initialize the node feature , and use as the input node feature vector and input it into the first layer of graph convolution; S2.42. For each node , determine the neighbor node set of node ; S2.43. Define a weight matrix and a bias term for each layer of convolutional operation; S2.44. In the first layer of graph convolutional operation, the node feature vector and the feature vectors of its neighbor nodes are aggregated through the graph convolution formula. The node type embedding vector and the real-time state factor are combined, and the median aggregation method is used to optimize the graph convolutional operation; S2.45. Based on the updated features of the previous layer, the node features continue to be aggregated with the feature vectors of its neighbor nodes until the predetermined number of layers is reached ; S2.46. After completing the convolution of all layers, the final feature representation of each node is obtained .

[0007] As a further improvement of this technical solution, in S2.44, the aggregation of the node feature vector and the feature vectors of its neighbor nodes through the graph convolution formula includes the following steps: S2.441. For the current node, traverse all its neighbor nodes, multiply the feature representation of each neighbor node in the th layer by the weight matrix of the current layer, and add the bias; S2.442. Normalize the weighted features of the neighbor nodes; S2.443. After summing the normalized features of all neighbor nodes, a new feature of the current node in the th layer is generated through a non-linear activation function.

[0008] As a further improvement of this technical solution, in S2.44, the combination of the node type embedding vector and the real-time state factor and the use of the median aggregation method to optimize the graph convolutional operation are as follows: To address the problem of heterogeneous node types in the power grid, assign node type embedding vectors to different node types , encode the node type features into the graph convolutional operation for optimization, and calculate the attention coefficient , dynamically quantify the importance weights of different neighbor nodes to the current node to obtain the attention coefficient ; Filter the low-confidence neighbors to obtain the filtered neighbor set : Adopt the median aggregation method to suppress the influence of outliers on feature aggregation to obtain the weighted median-processed neighbor node features : By combining the node type embedding vector with the real-time state factor , and using the median aggregation method to optimize the graph convolution formula after the graph convolutional operation.

[0009] As a further improvement of this technical solution, in S2.5, for the operation data of the power grid, the sliding window technique is used to extract the time series features in the operation data of the power grid, including the following steps: S2.51. Determine the size of the time window and set the sliding step size; S2.52. Starting from the starting point of the time series data, initialize the position of the first window, and successively move the window position through the sliding window method until the window slides to the last time point; S2.53. Extract key features within each window; S2.54. Construct a feature vector from the extracted features of each sliding window, and the feature vector is the time series feature.

[0010] As a further improvement of this technical solution, in S3, based on the fused spatio-temporal features, a long short-term memory network model is used for load forecasting, including the following steps: S3.1. Input the fused spatio-temporal features into the long short-term memory network model; S3.2. Process the time series data through the long short-term memory network layer to capture the time-dependent relationship of load changes; S3.3. Each long short-term memory network unit outputs a hidden layer state; S3.4. The output of the long short-term memory network is converted into a load forecast value through a fully connected layer; S3.5. Use the mean square error as the loss function to measure the difference between the forecast value and the true value of the long short-term memory network model, and train the long short-term memory network model; S3.6. Use the trained long short-term memory network model to forecast the load.

[0011] As a further improvement of this technical solution, in S4, based on the prediction results, an emergency resource scheduling optimization model is constructed through the dynamic programming method to generate the optimal resource scheduling path, including the following steps: S4.1. Define the state variables and the set of all positions of the resources ; S4.2. Define the decision variables and select the scheduling path in the current state t; S4.3. Define the state transition cost and the state transition equation, and introduce the task dependency graph and the hierarchical optimization strategy to optimize the state transition equation; S4.4. Set the initial state and the termination condition. The initial state is the starting point where the resources are located, and the termination state is that all tasks are completed; S4.5. Through reverse recursion, obtain the cumulative cost of each path reaching the final state, and select the path with the minimum cost among the cumulative costs of the paths as the optimal path; S4.6. Backtrack the optimal decision for each state to obtain the optimal resource scheduling path.

[0012] As a further improvement of this technical solution, in S4.3, the state transition equation is: ; where represents the minimum cumulative cost from the current time and node to the completion of all tasks; represents the node where the resource is at time ; represents the next node that the resource reaches at time ; represents the set of neighbor nodes that can be moved from the current node ; represents the comprehensive cost of moving from node to ; represents the minimum cumulative cost from the current time and node to the completion of all tasks; Regarding the sequence problem of power grid emergency resource scheduling, a task dependency graph is introduced in the state transition equation to optimize the scheduling path. Only when the prerequisite tasks of a node are completed, the node is added to the set : Define the task dependency matrix . If task depends on task , then ; Update the set of feasible nodes: , where represents all the tasks that have been completed when reaching state ; Divide the power grid area into sub-areas, optimize the state transition equation through a hierarchical optimization strategy, generate a coarse-grained path based on the key nodes of the sub-areas, refine the path within the sub-areas, with the global layer state variable being the sub-area ID and the local layer being the detailed node ID; Introduce a task dependency graph and a hierarchical optimization strategy to optimize the state transition equation.

[0013] As a further improvement of this technical solution, in S4.5, through reverse recursion, obtain the cumulative cost of each path reaching the final state, and select the path with the minimum cost among the cumulative costs of the paths, including the following steps: S4.51. Set the initial conditions for recursion; S4.52. Solve to obtain all the next states that can be transferred from the current state set ; S4.53. For each state in the set Calculate the total cost from transferring to and continuing to the end according to the state transition equation; S4.54. Among all the transfers, select the one with the minimum total cost as the value of the minimum cumulative cost, and record the corresponding optimal transfer state at the same time; S4.55. Repeat steps S4.52 to S4.54 until the optimal costs of all states at the initial moment are calculated.

[0014] Compared with the prior art, the beneficial effects of the present invention are: 1. In the grid emergency resource scheduling optimization method based on the multi-dimensional data fusion of source-network-load-storage, by integrating grid operation data, meteorological data, geographical information, and equipment status data, and using a graph neural network (GNN) for spatio-temporal feature fusion, this method can more accurately capture the complex structure and dynamic changes of the grid system. Further using a long short-term memory network (LSTM) model to predict the load can effectively capture the time-dependent relationship of load changes, thereby improving the accuracy of load prediction. Based on these prediction results, the emergency resource scheduling optimization model constructed by the dynamic programming (DP) method can generate the optimal resource scheduling path, greatly improving the efficiency and response speed of resource scheduling, especially being able to quickly respond in case of emergencies, reducing the power outage time, and enhancing the power supply reliability.

[0015] 2. In the grid emergency resource scheduling optimization method based on the multi-dimensional data fusion of source-network-load-storage, the heterogeneity of grid node types is considered. By assigning type embedding vectors to different node types and adjusting the attention coefficients in combination with real-time state factors, the model can better understand and process the characteristics of different types of nodes, enhancing the adaptability of the model to the dynamic changes of the grid. In addition, aiming at the problem that scheduling all resources at the global level will lead to too high computational complexity, a two-layer optimization strategy is adopted. First, at the global level, a coarse-grained path is generated based on the key nodes of sub-regions, and then the path is refined within the sub-regions. This method not only alleviates the curse of dimensionality problem but also speeds up the decision-making speed, making the emergency resource scheduling in large-scale grids more efficient and feasible. This greatly improves the ability to cope with complex grid environments and also ensures the effective utilization of resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 is the overall method flowchart of the present invention. Specific Embodiment

[0017] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0018] Embodiment: Please refer to Figure 1 As shown, this embodiment provides a power grid emergency resource scheduling optimization method based on the multi-dimensional data fusion of source, grid, load, and storage, including the following steps: S1. Deploy edge computing nodes at key power grid nodes (including substations, distribution terminals, and sensor networks) to collect real-time power grid operation data (current, voltage, power), meteorological data (wind speed, temperature, precipitation), geographical information data (topological structure of the power network, geographical location of equipment), and new energy output data (photovoltaic, wind power) on the source side to achieve real-time monitoring of all links of the source, grid, load, and storage; S2. Integrate power grid operation data, meteorological data, geographical information, and equipment status data, and use graph neural network (GNN) for spatio-temporal feature fusion; In this embodiment, the graph neural network (GNN) is a neural network model specifically designed to process graph-structured data. It learns the representations of nodes, edges, or the entire graph through information propagation of the nodes and edges of the graph; the power grid system itself is a complex network composed of different types of nodes (such as substations, distribution terminals, sensor networks, etc.) and edges (power line connections, communication networks, etc.). By treating the power grid as a graph and using the graph neural network (GNN), the complex spatial dependence relationships between nodes can be effectively captured, which is crucial for understanding the topological structure of the power grid and its dynamic changes; through the graph convolution operation and sliding window technology of the graph neural network, time series features are extracted, realizing the effective fusion of spatio-temporal features of the power grid system. This method not only considers the spatial characteristics of the power grid system but also combines its temporal variation laws, thereby improving the accuracy of the model in aspects such as future load prediction; Integrating power grid operation data, meteorological data, geographical information, and equipment status data, and using graph neural network (GNN) for spatio-temporal feature fusion includes the following steps: S2.1. Treat the power grid as a graph ; Among them, represents the nodes in the power grid (substations, distribution terminals, sensor networks, load points, etc.), represents the connection relationships between nodes (power line connections, communication networks, etc.); S2.2. For each node Merge into a comprehensive feature vector : ; Wherein represents power grid operation data (including source side output, load side demand, and charge and discharge power of energy storage, to characterize the collaborative operation characteristics of the source-grid-load-storage); represents meteorological data; represents geographical information; represents equipment status data (including status parameters such as remaining capacity and charge and discharge efficiency of energy storage equipment); S2.3. Initialize the features of each node of , integrate all features from different sources (power grid data, meteorological data, etc.) into the initial feature , as the input of the graph neural network; S2.4. Perform multi-layer graph convolution operations. The core of the graph neural network is graph convolution, which updates the features of nodes through a message passing mechanism. In the graph neural network, the features of nodes are affected by the information of their neighbor nodes in each layer of the network, forming the features of each node , including the spatial domain features, meteorological features, and geographical information features of the nodes; Wherein, performing multi-layer graph convolution operations includes the following steps: S2.41. Initialize the node features , and take as the input node feature vector and input it into the first layer of graph convolution; S2.42. For each node , determine the set of neighbor nodes of node ; S2.43. Define the weight matrix and bias term for each layer of convolution operation; S2.44. In the first layer of graph convolution operation, the node feature vector and the feature vectors of its neighbor nodes are aggregated through the graph convolution formula, combined with the node type embedding vector and the real-time state factor, and the median aggregation method is used to optimize the graph convolution operation; Furthermore, the aggregation of the node feature vector and the feature vectors of its neighbor nodes through the graph convolution formula includes the following steps: S2.441. For the current node, traverse all its neighbor nodes, multiply the feature representation of each neighbor node in the th layer by the weight matrix of the current layer, and superimpose the bias; S2.442. Normalize the weighted features of neighbor nodes, that is, multiply by a scaling factor to balance the influence of the number of neighbors of different nodes; S2.443. After summing the normalized features of all neighbor nodes, a new feature of the current node at the layer is generated through a non-linear activation function; The graph convolution formula is: ; Among them, represents the neighbor set of node ; represents the weight matrix of the layer; represents the bias term of the layer; represents the ReLU activation function; represents the index of the current node; represents the index of the neighbor node of node ; represents the feature representation of node after graph convolution in the layer, represents the neighbor nodes of node ; The power grid itself is a heterogeneous network, composed of different types of data nodes and connection relationships. Different types of nodes and edges may have different attributes and functions. In this case, if the differences in node types are not considered and graph convolution operations are directly performed, the model may ignore the essential differences between nodes, thus being unable to accurately capture the complex structure and dynamic changes of the power grid. Encoding node type features into graph convolution operations can enable the model to better adapt to the heterogeneous network characteristics of the power grid and improve the modeling ability of the power grid system; Regarding the problem of node type heterogeneity in the power grid (specifically manifested as the physical property differences between substations and sensors. Substations and sensors in the power grid have obvious different physical properties. Substations are power facilities in the power system that transform voltage, receive and distribute electric energy, control the flow of electric power, and adjust voltage. They are large in scale, complex in function, involving a large number of power equipment and complex electrical connections. While sensors are mainly used to monitor various physical quantities in the power grid, such as voltage, current, temperature, etc. Their functions are relatively single and their physical structures are relatively simple. Such differences in physical properties will lead to great differences in their roles and behavior patterns in the power grid; by assigning type embedding vectors to different node types, the feature information brought by these physical property differences can be encoded into graph convolution operations, enabling the model to better understand and process the characteristics of these different types of nodes), assign node type embedding vectors to different node types (substations, sensors) (The node type embedding vector includes the node function feature, the connection degree feature of the node, and the geographical location feature), encoding the node type feature into the graph convolution operation for optimization, and calculating the attention coefficient , dynamically quantifying the importance weights of different neighbor nodes for the current node (capturing the feature differences of different types of nodes through the embedding vector, enhancing the model's understanding ability of heterogeneous networks. In the power grid, the influence degrees of different neighbor nodes on the current node are different, and this influence degree may change with the operating state and time of the power grid. Through the attention mechanism, the model can dynamically adjust the importance weights of neighbor nodes according to the node type and the current state information, so as to more flexibly capture the relationships between nodes. For example, when a fault occurs, the substation adjacent to the fault node may have a greater impact on the current node, and the real-time monitoring data provided by sensors may also become more important. The attention coefficient enables the model to dynamically adjust the weights according to these situations, improving the model's adaptability to the dynamic changes of the power grid): ; Among them, denotes normalizing the calculation results of all neighbor nodes, denotes the weight matrix of the k-th layer, and [;] denotes vector concatenation; Introduce the real-time state factor (including equipment health and data confidence) to correct the attention coefficient, and this real-time state factor will affect the attention coefficient , used to correct the weight of the influence of neighbor nodes on the target node (through type embedding and state factor , distinguishing key nodes (such as substations in high-fault areas) from ordinary nodes), the introduction of the real-time state factor essentially deeply combines the dynamic operation characteristics of the power grid with the attention mechanism of graph convolution, enabling the model to dynamically adjust neighbor weights, adapt to the instantaneous changes of the power grid; distinguish key nodes, ensuring the feature dominance of core facilities; integrate dynamic and static information, enhancing the modeling ability of heterogeneous networks: ; Among them, denotes the corrected attention coefficient, considering the equipment health of the equipment or the confidence of the data; Filter out low-confidence neighbors (such as sensors with missing data). To reduce the impact of low-confidence nodes on the model, it is necessary to filter out those neighbors with poor equipment health or low data confidence: ; Among them, Denote the filtered neighbor set, which only contains those neighbor nodes with a higher real-time status factor ; Denote the filtering threshold; Adopt the median aggregation method (weighted median instead of mean) to suppress the influence of outliers on feature aggregation (in the power grid scenario, sensor data may have outliers due to equipment failures or communication interferences (such as mutated current / voltage values). If the mean aggregation is used, these outliers will significantly distort the feature representation and lead to misjudgments by the model. By weighting with the attention coefficient of the real-time status factor (such as data confidence), the median aggregation can not only suppress outliers but also retain the contributions of high-confidence neighbors, such as accidental outlier data that still exists after filtering; the weighted median can more flexibly reflect the priorities of key nodes (such as substations with high health levels) while reducing the interference of low-confidence nodes): ; Among them, Denote the weighted median operation; Denote the feature of neighbor nodes after weighted median processing; To sum up, by combining the node type embedding vector and the real-time status factor , and adopting the median aggregation method to optimize the graph convolution formula after graph convolution operation is: ; Among them, Denote the feature of the optimized node; Denote the multi-layer perceptron, which is used to further process the feature of node ; Denote the weight coefficient, which is used to control the contribution ratio of the node's own feature (the result after MLP processing) in the overall feature update.

[0019] S2.45. Based on the updated features of the previous layer, the node features continue to be aggregated with the feature vectors of their neighbor nodes. At each layer, the node features are updated according to the features of the neighbor nodes, continuously transmitting information until the predetermined number of layers is reached. For each additional layer, the node will be able to obtain information from more distant neighbors, thereby capturing a larger range of spatial dependencies; S2.46. After completing the convolution of all layers, the final feature representation of each node is obtained .

[0020] S2.5. For the operation data of the power grid, use the sliding window technique to extract the time series features in the operation data of the power grid; Among them, for the operation data of the power grid, the sliding window technique is used to extract the time series features in the operation data of the power grid, including the following steps: S2.51. Determine the size of the time window (i.e., the length of the time period processed each time), and set the sliding step (i.e., the distance the window moves each time); S2.52. Starting from the starting point of the time series data, initialize the position of the first window, and by the sliding window method, move the window position sequentially until the window slides to the last available time point; S2.53. In each window, extract key features, including statistical features and frequency domain features; S2.54. Construct a feature vector from the extracted features of each sliding window, that is, the time series features; S2.6. Combine (concatenate) the time series features obtained from time series analysis with the spatial domain features obtained through the graph neural network to form spatio-temporal features. The graph convolution operation aggregates the time series features of neighboring nodes in the spatio-temporal domain to achieve the integration of spatio-temporal features.

[0021] S3. Based on the fused spatio-temporal features, use the long short-term memory network (LSTM) model for load forecasting; In this embodiment, LSTM is a special recurrent neural network (RNN), which is particularly good at processing and predicting time series data. It can effectively capture the long-term dependencies in the time series, which is particularly important for load forecasting because the power load is usually affected by historical load patterns, seasonal changes, periodic factors such as weekdays / weekends, etc.; by fusing the spatio-temporal features of the power grid (including geographical location information, the status of neighboring nodes, etc.), LSTM can use the spatial features extracted by the graph neural network (GNN) to enhance the performance of the prediction model. This method not only considers the changing trend in the time dimension but also integrates the mutual influence between different geographical locations, making the prediction more accurate and comprehensive; Based on the fused spatio-temporal features, use the long short-term memory network (LSTM) model for load forecasting, including the following steps: S3.1. Input the fused spatio-temporal features into the long short-term memory network model. These spatio-temporal features are a combination of the operation data of the power grid, meteorological data, geographical information, and equipment status information; S3.2. Process the time series data through the long short-term memory network layer to capture the time dependencies of the load changes. In this layer, LSTM will have two main gating mechanisms: the input gate and the forget gate, which are used to control the storage and forgetting of information and can effectively capture long-term dependencies; S3.3. Each long short-term memory network unit outputs a hidden layer state, which captures the information of the historical input sequence; S3.4. The output of the long short-term memory network is converted into a load prediction value through a fully connected layer (i.e., a multi-layer perceptron). Here, the output layer will predict the value of the power grid load; S3.5. Use the mean squared error (MSE) as the loss function to measure the difference between the predicted value and the true value of the long short-term memory network model, and train the long short-term memory network model; The mean squared error is: ; where represents the mean squared error; represents the number of data points for prediction; represents the th true load value of a data point; represents the th load value predicted by the long short-term memory network model for a data point.

[0022] S3.6. Use the trained long short-term memory network model to predict the load.

[0023] S4. Based on the prediction results, construct an emergency resource scheduling optimization model through the dynamic programming (DP) method to generate the optimal resource scheduling path; In this embodiment, dynamic programming decomposes a complex problem into a series of interrelated sub-problems and uses the backward recursion method to solve it, and can obtain the global optimal solution from the initial state to the target state. This is particularly important for emergency resource scheduling because it involves multiple constraints and variables, and directly solving may lead to local optimality; emergency resource scheduling is usually a complex process with multiple decision points, and each decision will affect subsequent choices. Dynamic programming can effectively handle this type of multi-stage decision problem, describe the relationship between different decisions through the state transition equation, and thus achieve step-by-step optimization; determine the overall optimal path by backtracking the optimal decision of each state. This process can not only obtain the final solution, but also allow checking and verification of the intermediate steps to ensure that each decision is made based on the current best information; Based on the prediction results, construct an emergency resource scheduling optimization model through the dynamic programming (DP) method to generate the optimal resource scheduling path, including the following steps: S4.1. Define the state variable , which represents the location where a certain resource (such as a repair person or equipment) is located, the tasks that have been completed, and the set of all locations of the resources ; S4.2. Define decision variables (decision variables refer to the specific paths chosen to transfer from the current state to the next state, that is, the decisions to move from the current node to the next node. Each decision determines the next location where the resources will go), select the scheduling path at the current state t. The decision at each moment is to select the next scheduling path, that is, from the current node to the next node; S4.3. Define the state transition cost and the state transition equation. The core of dynamic programming is the state transition equation, which describes the cost of transferring from one state to another. The state transition cost refers to the cost or resources consumed for transferring from one state to another. The state transition equation is used to describe how to calculate the best way to reach the next state based on the current state and determine the corresponding transition cost; Furthermore, the state transition equation is: ; ; where, represents the minimum cumulative cost from the current time and node to the completion of all tasks; represents the node (substation, maintenance center, etc.) where the resource is located at time ; represents the next node reached by the resource at time ; represents the set of neighbor nodes that can be moved from the current node (i.e., the optional next locations); represents the comprehensive cost (including time, resource consumption, traffic impact, etc.) of moving from node to ; represents the minimum cumulative cost (recursive calculation) from the current time and node to the completion of all tasks; represents the time cost (distance divided by speed) of moving from node to ; represents the resource consumption (manpower, equipment loss); represents the road congestion level; represents the weight coefficient of the time cost; represents the weight coefficient of the resource consumption; represents the weight coefficient of the road congestion level; There is often an inherent logical sequence among tasks in power grid emergency handling. For example, when repairing a transmission line fault, it may be necessary to first detect and locate the fault point before arranging maintenance personnel and equipment for the repair work; after the repair is completed, testing and acceptance are also required to ensure the line resumes normal operation. The task dependency graph (DAG) can accurately reflect the dependency relationships among these tasks. By introducing DAG constraints for the scheduling path, it can ensure that resource scheduling proceeds in a reasonable task order, avoiding unreasonable scheduling that violates the task logic. If the sequence of tasks is not considered, it may lead to resources being allocated to a task at an inappropriate time, resulting in resource conflicts and waste. For example, dispatching maintenance personnel and equipment to the site without completing the fault detection may cause the repair work to not proceed smoothly due to the inability to accurately understand the fault situation, wasting human and material resources. After introducing DAG constraints, it can ensure that only when the prerequisite tasks are completed, the resources required for subsequent tasks are scheduled, improving the utilization efficiency of resources; Regarding the sequence problem of power grid emergency resource scheduling, the task dependency graph (DAG) constraint scheduling path is introduced into the state transition equation for optimization. Only when the prerequisite tasks of a node are completed, the node is added to the set , which can make the scheduling process more targeted, avoid unnecessary calculations and attempts, and speed up the decision-making speed of emergency resource scheduling: Define the task dependency matrix , if task depends on task , then ; Among them, the task dependency matrix is to clarify the dependency relationships among various tasks in power grid emergency resource scheduling, so as to optimize the scheduling path. First, it is necessary to determine all the tasks involved in the power grid emergency resource scheduling process. These tasks can be different maintenance work, equipment detection, power supply restoration, etc. operations. Number these tasks. If there are a total of tasks, they are respectively marked as ; conduct a detailed analysis of each task to determine the sequence and dependency relationships among them. For example, when performing power equipment maintenance, it may be necessary to first conduct equipment detection before carrying out specific maintenance work, which means that the maintenance task depends on the detection task; construct a matrix , the rows and columns of the matrix respectively correspond to each task; the element in the matrix represents the dependency relationship between task and task , represents the task number corresponding to the row, Indicates the task number corresponding to the column; among them, ; According to the analysis result of the task dependency relationship, assign values to the elements in the matrix : If task depends on task , then let , which means that task must be completed before task starts; If task does not depend on task , then let ; Update the set of feasible nodes: , among which, represents all the tasks that have been completed when reaching state ; When scheduling all resources at the global level, the state space will expand rapidly as the scale of the power grid increases. The power grid contains a large number of nodes (such as substations, maintenance centers, etc.) and resources (such as repair personnel, equipment), and each node and resource has multiple possible states and combinations, which makes the dimension of the state variables very high. If the entire power grid is directly scheduled uniformly, the amount of calculation will increase exponentially, resulting in extremely high computational complexity and even becoming infeasible in practical applications. After dividing the power grid into sub-regions, the scale of the state space of each sub-region is relatively small, and the computational complexity will also be reduced accordingly, thus alleviating the curse of dimensionality problem; Global scheduling needs to consider all possible scheduling paths of all resources in the entire power grid, and the search space is extremely large. This means that when solving the optimal scheduling path, a large number of possible solutions need to be traversed, and the computational cost is huge and time-consuming. Through two-layer optimization, first generate a coarse-grained path based on the key nodes of the sub-region at the global level, narrow the search scope, and reduce unnecessary calculations; then refine the path within the sub-region to further improve the accuracy of scheduling, while avoiding the high complexity brought by a comprehensive search of the entire power grid; For the problem that scheduling all resources at the global level will lead to too high computational complexity, divide the power grid area into sub-regions, optimize the state transition equation through a hierarchical optimization strategy, generate a coarse-grained path based on the key nodes of the sub-region, and refine the path within the sub-region. The global layer state variable is the sub-region ID, and the local layer is the detailed node ID, to solve the curse of dimensionality problem; In summary, introducing the task dependency graph and the hierarchical optimization strategy to optimize the state transition equation is: ; Among them, represents the optimized minimum cumulative cost; represents the state variable at the sub-region level for hierarchical optimization; Denotes the adjustment factor, which controls the weights of the global layer and the local layer to optimize the computational efficiency. The value range is [0, 1], and it is determined by grid search; Denotes starting from the next time point and the state variables at the sub-region level to the minimum cumulative cost until all tasks are completed.

[0024] S4.4. Set the initial state and termination conditions. The initial state is the starting point where the resources are located, and the termination state is when all tasks are completed; S4.5. Through backward recursion, obtain the cumulative cost of each path reaching the final state, and select the path with the minimum cost among the cumulative costs of the paths as the optimal path; Among them, through backward recursion, obtaining the cumulative cost of each path reaching the final state and selecting the path with the minimum cost among the cumulative costs of the paths as the optimal path includes the following steps: S4.51. Set the initial conditions for recursion, that is, determine the optimal costs of all possible states at the final moment; S4.52. Solve to obtain all the next states that can be transferred from the current state set ; S4.53. For each state in the set calculate the total cost of transferring from to and continuing to the end point, including the direct transfer cost plus the minimum cost from to the end point in the future; S4.54. Among all the transfers, select the one with the minimum total cost as the value of the minimum cumulative cost, and record the corresponding optimal transfer state at the same time; S4.55. Repeat steps S4.52 to S4.54 until the optimal costs of all states at the initial moment are calculated. At this time, the optimal cost of the initial state is the minimum scheduling cost from the starting point to the end point. In this way, the optimal path to reach the target state and its corresponding minimum cost can be found from any starting point within the entire time range. This is actually constructing a global optimal strategy to ensure that the resources reach the predetermined target position at the lowest cost during the entire scheduling process; S4.6. Backtrack the optimal decisions of each state to obtain the optimal resource scheduling path.

[0025] The foregoing has shown and described the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments, and the above embodiments and the descriptions in the specification are only preferred examples of the present invention and are not used to limit the present invention. Without departing from the spirit and scope of the present invention, the present invention will have various changes and improvements, and these changes and improvements fall within the scope of the present invention claimed.

Claims

1. A power grid emergency resource dispatch optimization method based on source, grid, load and storage multi-dimensional data fusion, characterized in that: The following steps are involved: S1. Deploy edge computing nodes at key nodes of the power grid to collect real-time power grid operation data, meteorological data, and geographic information data; S2. Integrate power grid operation data, meteorological data, geographic information and equipment status data, use graph neural network to fuse spatiotemporal features, combine node type embedding vectors with real-time status factors, and use median aggregation method to optimize the process of spatiotemporal feature fusion of graph neural network; S3, based on the fused spatiotemporal characteristics, the long short-term memory network model is used for load forecasting; S4. Based on the prediction results, an emergency resource scheduling optimization model is constructed through dynamic programming method to generate the optimal resource scheduling path.

2. The power grid emergency resource dispatch optimization method based on source-grid-load-storage multi-dimensional data fusion according to claim 1 is characterized by: In S2, the power grid operation data, meteorological data, geographic information and equipment status data are integrated, and the spatiotemporal feature fusion is performed using a graph neural network, including the following steps: S2.

1. Consider the power grid as a graph ; in, represents a node in the power grid, Indicates the connection relationship between nodes; S2.

2. For each node Combined into a comprehensive feature vector ; S2.

3. Initialize each node Features , integrating features from all different sources into initial features , as the input of the graph neural network; S2.

4. Perform multi-layer graph convolution operations to form the features of each node , Contains spatial domain features, meteorological features and geographic information features of nodes; S2.

5. For the operation data of the power grid, the sliding window technology is used to extract the time series characteristics in the operation data of the power grid; S2.

6. Combine the time series features obtained from time series analysis with the spatial domain features obtained through graph neural networks to form spatiotemporal features.

3. The power grid emergency resource dispatch optimization method based on source-grid-load-storage multi-dimensional data fusion according to claim 2 is characterized by: In S2.4, a multi-layer graph convolution operation is performed, including the following steps: S2.

41. Initialize node features ,Will The node feature vector as input is passed into the first layer of graph convolution; S2.

42. For each node , determine the node The set of neighbor nodes of S2.43, define the weight matrix and bias term for each layer of convolution operation; S2.

44. In the first layer of graph convolution operation, the feature vector of a node is aggregated with the feature vectors of its neighboring nodes through the graph convolution formula. The graph convolution operation is optimized by combining the node type embedding vector and the real-time state factor and using the median aggregation method. S2.

45. Based on the updated features of the previous layer, the node features continue to aggregate with the feature vectors of its neighboring nodes until the predetermined number of layers is reached. ; S2.

46. After completing the convolution of all layers, the final feature representation of each node is obtained .

4. The power grid emergency resource dispatch optimization method based on source-grid-load-storage multi-dimensional data fusion according to claim 3 is characterized by: In S2.44, the node feature vector and the feature vectors of its neighboring nodes are aggregated through a graph convolution formula, including the following steps: S2.

441. For the current node, traverse all its neighboring nodes and place each neighboring node in the The feature representation of the layer is multiplied by the weight matrix of the current layer and the bias is superimposed; S2.442, normalizing the weighted features of neighboring nodes; S2.443, after the normalized features of all neighbor nodes are summed, the current node is generated through a nonlinear activation function. New features of the layer.

5. The power grid emergency resource dispatch optimization method based on source-grid-load-storage multi-dimensional data fusion according to claim 3 is characterized by: In S2.44, the graph convolution operation is optimized by combining the node type embedding vector and the real-time state factor and using the median aggregation method as follows: To address the heterogeneity of power grid node types, node type embedding vectors are assigned to different node types. , encode the node type features into the graph convolution operation for optimization, and calculate the attention coefficient , dynamically quantify the importance weights of different neighbor nodes to the current node, and obtain the attention coefficient ; Filter low-confidence neighbors to obtain a filtered neighbor set : The median aggregation method is used to suppress the influence of outliers on feature aggregation and obtain the neighbor node features after weighted median processing. : By combining the node type embedding vector With real-time status factor , and the median aggregation method is used to optimize the graph convolution formula after the graph convolution operation.

6. The power grid emergency resource dispatch optimization method based on source-grid-load-storage multi-dimensional data fusion according to claim 2 is characterized by: In S2.5, for the operation data of the power grid, a sliding window technique is used to extract time series features in the operation data of the power grid, including the following steps: S2.51, determine the time window size and set the sliding step size; S2.

52. Starting from the starting point of the time series data, initialize the position of the first window, and use the sliding window method to move the window position in sequence until the window slides to the last time point; S2.53, in each window, extract key features; S2.

54. The features extracted from each sliding window are used to form a feature vector, which is the time series feature.

7. The power grid emergency resource dispatch optimization method based on source-grid-load-storage multi-dimensional data fusion according to claim 1 is characterized by: In S3, based on the fused spatiotemporal features, a long short-term memory network model is used to perform load forecasting, including the following steps: S3.1, input the fused spatiotemporal features into the long short-term memory network model; S3.2, process time series data through long short-term memory network layer to capture the time dependency of load changes; S3.3, each LSTM unit outputs a hidden state; S3.4, the output of the long short-term memory network is converted into load forecast value through the fully connected layer; S3.

5. Use the mean square error as the loss function to measure the difference between the predicted value and the true value of the long short-term memory network model, and train the long short-term memory network model; S3.

6. Use the trained long short-term memory network model to predict the load.

8. The power grid emergency resource dispatch optimization method based on source-grid-load-storage multi-dimensional data fusion according to claim 1 is characterized by: In S4, based on the prediction results, an emergency resource scheduling optimization model is constructed by a dynamic programming method to generate an optimal resource scheduling path, including the following steps: S4.

1. Define state variables and all locations of resources ; S4.2, define decision variables and select the scheduling path under the current state t; S4.3, define the state transfer cost and state transfer equation, and introduce the task dependency graph and hierarchical optimization strategy to optimize the state transfer equation; S4.4, set the initial state and termination conditions. The initial state is the starting point where the resources are located, and the termination state is when all tasks are completed; S4.

5. Obtain the cumulative cost of each path to the final state through reverse recursion, and select the path with the smallest cost as the optimal path among the cumulative costs of the paths; S4.

6. Backtrack the optimal decision for each state to obtain the optimal resource scheduling path.

9. The power grid emergency resource dispatch optimization method based on source-grid-load-storage multi-dimensional data fusion according to claim 8 is characterized by: In S4.3, the state transfer equation is: ； in, Indicates from the current time and nodes The minimum cumulative cost from the start to the completion of all tasks; Indicates that the resource is at time The node where it is located; Indicates that the resource is at time The next node to reach; Indicates that from the current node The set of neighbor nodes that can be moved to; Represents a slave node Move to The overall cost of Indicates from the current time and nodes The minimum cumulative cost from the start to the completion of all tasks; In order to solve the problem of the order of emergency resource dispatch in power grid, the task dependency graph is introduced into the state transfer equation to constrain the dispatch path for optimization. The node is added to the set only when the node's predecessor task is completed. : Defining the task dependency matrix , if the task Dependent tasks ,but ; Update the set of feasible nodes: ,in, Indicates arrival status All completed tasks; The power grid area is divided into sub-areas, and the state transfer equation is optimized through a hierarchical optimization strategy. A coarse-grained path is generated based on the key nodes of the sub-area, and the path is refined within the sub-area. The state variable of the global layer is the sub-area ID, and the local layer is the detailed node ID. Task dependency graph and hierarchical optimization strategy are introduced to optimize the state transfer equation.

10. The power grid emergency resource dispatch optimization method based on source-grid-load-storage multi-dimensional data fusion according to claim 9 is characterized in that: In S4.5, the cumulative cost of each path to the final state is obtained by reverse recursion, and the path with the smallest cost is selected as the optimal path among the cumulative costs of the paths, including the following steps: S4.51, set the initial conditions for recursion; S4.52, solve to get from the current state All next states that can be transferred to Collection ; S4.

53. For the collection Each state in According to the state transfer equation, the Transfer to and continue to the total cost of the end point; S4.

54. Among all the transfers, select the one with the smallest total cost as the value with the smallest cumulative cost, and record the corresponding optimal transfer state; S4.55, repeat steps S4.52 to S4.54 until the initial time is calculated The optimal cost for all states.

Images