Power grid source-load prediction method and device based on space-time diagram neural network

Abstract

The invention provides a power grid source-load prediction method and device based on a space-time diagram neural network, and relates to the technical field of power grids. According to the method, the medium and long term planning scheme is analyzed into the graph evolution operation instruction and the dynamic planning graph sequence is generated, so that complete mathematical description of the power grid growth process is realized, the prediction model can accurately accord with a time sequence path of power grid planning construction, and the problem that a static model is not matched with dynamic planning is solved. And then, based on the dynamic planning graph sequence, combining with a mixed space-time diagram neural network, carrying out spatial feature aggregation and time sequence deduction, considering spatial constraints and a source load long-term evolution law brought by power grid topology change, and outputting high-quality and high-reliability source load prediction data. The method solves the problem that the static power grid topology is disjointed from the reality when being used for source load prediction, and improves the accuracy of power grid source load prediction.

Description

Technical Field

[0001] This invention relates to the field of power grid technology, and in particular to a power grid source-load prediction method and apparatus based on spatiotemporal graph neural networks. Background Technology

[0002] Theoretical line loss calculation for power grids is a key basis for evaluating the economic operation level of power grids, optimizing network structure, and formulating energy-saving and consumption-reducing strategies. Among them, accurate prediction of power sources (generation side) and loads (consumption side) in the power grid, i.e., source-load prediction, is an important prerequisite for accurate theoretical line loss calculation, especially for forward-looking analysis in view of power grid development planning.

[0003] Currently, most existing source-load forecasting methods are based on historical operational data and employ time series models (such as ARIMA) or machine learning models (such as recurrent neural networks (RNN) and long short-term memory networks (LSTM) for modeling. These traditional models reveal a fundamental technical flaw in practice: they are typically built upon a fixed and unchanging power grid topology. The models assume that when predicting future source-load distribution, the grid's connectivity, node and line composition are completely consistent with its historical and current state.

[0004] However, in actual power grid planning and construction practice, the power grid is a complex system that is constantly evolving dynamically. Long-term power grid planning schemes clearly guide future changes in the power grid structure, such as: building new distributed photovoltaic power stations, wind farms, and other new energy nodes; adding or upgrading transmission lines to strengthen regional interconnection; expanding substations; and the migration of load centers along with urban development. These changes will profoundly alter the distribution paths of power flows, electrical distances, and the spatiotemporal coupling characteristics of power sources and loads.

[0005] Therefore, when using traditional static topology-based models to predict a dynamically evolving future power grid, the prediction results will inevitably be severely out of sync with the actual state of the future power grid. This out-of-sync directly leads to a significant decrease in the accuracy of source-load predictions for the planned state. Summary of the Invention

[0006] This invention provides a power grid source-load prediction method and apparatus based on a spatiotemporal graph neural network, which solves the problem of disconnect between the prediction method and the actual situation when using static power grid topology, and improves the accuracy of power grid source-load prediction.

[0007] In a first aspect, the present invention provides a power grid source-load prediction method based on a spatiotemporal graph neural network. The method includes: acquiring a medium- and long-term planning scheme for the power grid and current power grid topology data; performing power grid structure change analysis based on the medium- and long-term planning scheme to obtain multiple graph evolution operation instructions; performing sequential deduction based on the multiple graph evolution operation instructions and the current power grid topology data to generate a dynamic planning graph sequence of the power grid topology from the current basic state to the future planned state, wherein the dynamic planning graph sequence is used to characterize the dynamic evolution process of the power grid topology over time; and performing spatial feature aggregation and temporal deduction based on the dynamic planning graph sequence and a pre-constructed hybrid spatiotemporal graph neural network model to obtain predicted source-load power values ​​and key operational data deduction values ​​for each node of the power grid within a future target time period.

[0008] Secondly, embodiments of the present invention provide a power grid source-load prediction device based on a spatiotemporal graph neural network. The device includes a communication module and a processing module. The communication module is used to acquire medium- and long-term planning schemes for the power grid and current power grid topology data. The processing module is used to perform power grid structure change analysis based on the medium- and long-term planning scheme to obtain multiple graph evolution operation instructions. Based on the multiple graph evolution operation instructions and the current power grid topology data, it performs serialized deduction to generate a dynamic planning graph sequence of the power grid topology from the current basic state to the future planning state. The dynamic planning graph sequence is used to characterize the dynamic evolution process of the power grid topology over time. Based on the dynamic planning graph sequence and a pre-constructed hybrid spatiotemporal graph neural network model, it performs spatial feature aggregation and temporal deduction to obtain the predicted source-load power values ​​and key operational data deduction values ​​of each node in the power grid within the future target time period.

[0009] Thirdly, embodiments of the present invention provide an electronic device including a memory and a processor. The memory stores a computer program, and the processor is configured to call and run the computer program stored in the memory to perform the steps of the method as described in the first aspect and any possible implementation thereof.

[0010] Fourthly, embodiments of the present invention provide a computer-readable storage medium storing a computer program, characterized in that, when the computer program is executed by a processor, it implements the steps of the method as described in the first aspect and any possible implementation thereof.

[0011] This invention provides a power grid source-load prediction method and apparatus based on a spatiotemporal graph neural network. By parsing medium- and long-term planning schemes into graph evolution operation instructions and generating a dynamic programming graph sequence, this invention achieves a complete mathematical description of the power grid growth process. This allows the prediction model to accurately match the temporal path of power grid planning and construction, solving the problem of mismatch between static models and dynamic planning. Subsequently, based on the dynamic programming graph sequence and combined with a hybrid spatiotemporal graph neural network, spatial feature aggregation and temporal extrapolation are performed. Considering the spatial constraints brought about by power grid topology changes and the long-term evolution law of source and load, high-quality and highly reliable source-load prediction data is output. This solves the problem of disconnect between static power grid topology and reality in source-load prediction, improving the accuracy of power grid source-load prediction. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 This is a flowchart illustrating a power grid source-load prediction method based on a spatiotemporal graph neural network provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of a power grid source-load prediction device based on a spatiotemporal graph neural network provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0014] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of the invention. However, those skilled in the art will understand that the invention can be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods are omitted so as not to obscure the description of the invention with unnecessary detail.

[0015] In the description of this invention, unless otherwise stated, " / " means "or". For example, A / B can mean A or B. The term "and / or" in this document is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, and B alone. Furthermore, "at least one" and "more than one" refer to two or more. The terms "first," "second," etc., do not limit the quantity or order of execution, and "first," "second," etc., do not necessarily imply differences.

[0016] In the embodiments of this application, the terms "exemplary" or "for example" are used to indicate that something is an example, illustration, or description. Any embodiment or design that is described as "exemplary" or "for example" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design. Specifically, the use of terms such as "exemplary" or "for example" is intended to present the relevant concepts in a specific manner to facilitate understanding.

[0017] Furthermore, the terms "comprising" and "having," and any variations thereof, used in the description of this application are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or modules is not limited to the steps or modules listed, but may optionally include other steps or modules not listed, or may optionally include other steps or modules inherent to such process, method, product, or device.

[0018] To make the objectives, technical solutions, and advantages of the present invention clearer, the following description will be provided in conjunction with the accompanying drawings and specific embodiments.

[0019] like Figure 1 As shown, this embodiment of the invention provides a power grid source-load prediction method based on a spatiotemporal graph neural network. The method includes steps S101-S104.

[0020] S101. Obtain the medium- and long-term planning schemes for the power grid and the current power grid topology data.

[0021] In some embodiments, medium- to long-term planning schemes include a mixture of structured and unstructured data, including the geographical locations and installed capacity of new power plants, planned transmission corridors and substations, obsolete equipment expected to be decommissioned, and the projected growth distribution of load centers.

[0022] In some embodiments, the current power grid topology data represents the complete power grid connectivity, including node types (generator nodes, load nodes, tie nodes, etc.), node basic parameters (rated voltage, capacity reference value), branch connectivity (lines, transformers), and branch parameters (resistance, reactance, susceptance). S102. Based on the medium- and long-term planning scheme, perform power grid structure change analysis to obtain multiple graph evolution operation instructions.

[0023] In some embodiments, the power grid structure change resolution process is essentially a crucial bridge in converting planning documents into machine-executable instructions. This process uses predefined resolution rules to identify descriptions of power grid structure changes in the planning and standardizes them into computer-processable instruction units.

[0024] As one possible implementation, step S102 can be specifically implemented as steps S1021-S1024.

[0025] S1021. Based on the text and drawings of the medium- and long-term planning scheme, identify and extract the key planning elements for power grid structure changes.

[0026] In some embodiments, the key planning elements include power generation equipment, transformer equipment, load nodes to be added or decommissioned, and transmission lines to be newly built or upgraded.

[0027] For example, embodiments of the present invention can employ a hybrid parsing method based on rules and pattern matching. For the text content of the planning scheme, key information involving changes to the power grid structure is automatically located and extracted using predefined power grid equipment named entity recognition rules, including core attributes such as equipment name, equipment type, capacity specifications, and spatial location. For the planning drawings, the symbols of newly added or decommissioned electrical equipment and their connection relationships are identified by parsing their standardized element layer information.

[0028] S1022. For each key planning element, analyze the graph evolution operation instruction type and map it to a graph evolution atomic operation instruction.

[0029] The graph evolution atomic operation instructions include adding nodes, deleting nodes, adding edges, deleting edges, modifying node attributes, and modifying edge attributes.

[0030] For example, embodiments of the present invention can perform precise conversion based on a pre-established element type-instruction type mapping rule base. For instance, when a planning element of a newly added photovoltaic power station is identified, it will be mapped sequentially to an add node instruction (used to create a grid-connected node for the power station) and a modify node attribute instruction (used to set the node type to photovoltaic power station and enter the planned capacity). For line modification elements, it may involve a combination of delete edge (removing the old line) and add edge (creating a new line) instruction mappings.

[0031] S1023. Set execution parameters for each graph evolution atom operation instruction to obtain the operation instructions after setting the parameters.

[0032] The execution parameters of node-related instructions in the operation instructions include the node's unique identifier, node type, and planned capacity; the execution parameters of edge-related instructions in the operation instructions include the source node and target node identifier pair, line type, and planned impedance.

[0033] For example, embodiments of the present invention can employ a hierarchical parameter generation strategy. For node-related instructions, their unique identifiers are automatically generated according to the expected location and type of the equipment in the power grid, following the power grid coding standards; node types are labeled according to standard classification systems such as power generation equipment, substation equipment, and load nodes; and planned capacity is directly read from the structured data fields of the planning scheme or parsed and extracted from unstructured descriptions. For edge-related instructions, their node identifiers are automatically generated by associating the starting and ending devices connected through parsing the lines; and line models and planned impedances are matched and assigned values ​​based on the standard line parameter library or the technical specifications in the planning scheme.

[0034] S1024. Based on the project commissioning time sequence and grid connection logic dependency relationship in the medium and long-term planning scheme, the operation instructions after setting parameters are sorted by time sequence and logically grouped to generate multiple graph evolution operation instructions with clear execution order.

[0035] For example, embodiments of the present invention can assign precise timestamps to each operation instruction based on the project commissioning time explicitly specified in the planning scheme. For projects with complex time dependencies, a topology sorting algorithm based on the critical path is used to ensure that instructions with logical dependencies (such as the requirement to build a substation before connecting to the line) are executed in the correct order. Instructions with the same execution timestamp are automatically grouped into the same batch of operation instruction groups. For phased construction projects spanning multiple years, the instruction groups are further divided into different planning phase sets based on planning phase milestones, ultimately forming a graph evolution operation instruction sequence with clear time dimensions and logical hierarchies.

[0036] S103. Based on the multiple graph evolution operation instructions and the current power grid topology data, perform serialization deduction to generate a dynamic planning graph sequence of the power grid topology from the current basic state to the future planning state.

[0037] In this embodiment of the application, the dynamic programming graph sequence is used to characterize the dynamic evolution of the power grid topology over time.

[0038] In some embodiments, the sequential deduction process simulates the growth process of the power grid. Starting from the current power grid, each graph evolution operation instruction is executed sequentially according to the planned timeline and logical order. Each time an instruction or group of instructions is executed, a "snapshot" representing the state of the power grid at that point in time is generated. All these "snapshots" are arranged in chronological order, ultimately forming a complete dynamic graph sequence that describes the gradual evolution of the power grid from its current form to its planned future form.

[0039] As one possible implementation, step S103 can be specifically implemented as steps S1031-S1035.

[0040] S1031. Based on the current power grid topology data, construct an initial state diagram as the starting point for the deduction.

[0041] For example, embodiments of the present invention can perform full-element digital modeling based on current power grid topology data. The node profile data of the current power grid is analyzed, and each substation, power plant, load node, etc., is abstracted as a vertex in a graph structure. Each vertex is assigned actual electrical attributes, including node type, current operating capacity, voltage level, and geographical location code. Simultaneously, the power grid line connection relationships are analyzed, and transmission lines, transformers, and other connecting equipment are abstracted as edges in the graph. Each edge is configured with its actual electrical parameters, including physical characteristics such as line impedance, admittance, rated capacity, and length.

[0042] S1032. Based on the aforementioned medium- and long-term planning scheme, construct a target planning map as the focus of the simulation.

[0043] S1033. Sort the multiple graph evolution operation instructions according to the project commissioning time order, and group the operation instructions belonging to the same timestamp into the same instruction group to obtain the time sequence instruction group sequence.

[0044] For example, embodiments of the present invention may employ a multi-level timeline alignment strategy. For each graph evolution operation instruction, its corresponding planning project implementation time information is parsed, and a timestamp accurate to the month and year is established. For complex projects constructed in phases, they are decomposed into multiple sub-instructions with continuous time dependencies. Based on power grid construction logic, the dependency chain between instructions is identified, ensuring that construction instructions for basic supporting facilities take precedence over subsequent instructions that depend on them. Finally, a time window sliding algorithm is used to automatically aggregate instructions within the same construction cycle (such as the same quarter or the same year) that have no logical conflicts into instruction groups, forming an instruction execution sequence with temporal coherence and logical rationality.

[0045] S1034. Starting from the initial state diagram, execute the timing instruction group sequence in chronological order to obtain multiple intermediate state diagrams.

[0046] Each time an instruction group is successfully executed, the power grid topology diagram is updated, and the updated power grid topology diagram is used as the intermediate state diagram after the execution of that instruction group.

[0047] For example, embodiments of the present invention can implement an incremental graph structure evolution mechanism based on instruction groups. Taking the previous state graph as input, all operation instructions in the current instruction group are executed sequentially: for node addition instructions, a new vertex is created in the graph and its attributes are fully configured; for node deletion instructions, the specified vertex and its associated edges are removed; for edge operation instructions, the connection relationships between vertices are adjusted accordingly. After each instruction group is executed, a graph structure consistency verification process is initiated to verify topological characteristics such as node degree distribution and network connectivity, ensuring that the evolved graph structure conforms to the physical connection constraints of the power grid. The verified state is saved as a new intermediate state graph, forming a state sequence of power grid topology evolution.

[0048] S1035. Based on the initial state diagram, the multiple intermediate state diagrams, and the target planning diagram, they are combined in chronological order to form a dynamic planning diagram sequence.

[0049] For example, embodiments of the present invention can employ a spatiotemporally aligned sequence construction method. The initial state diagram, each intermediate state diagram, and the target planning diagram are arranged in strict temporal order according to their corresponding timestamps to construct a complete timeline. For each state diagram, in addition to saving its topology structure, the planning stage identifier, construction progress percentage, and key time node metadata corresponding to that state are also recorded. Through this organization, the resulting dynamic planning diagram sequence not only completely records the evolution path of the power grid topology but also retains the temporal context information of each evolution stage.

[0050] S104. Based on the dynamic programming graph sequence and the pre-built hybrid spatiotemporal graph neural network model, perform spatial feature aggregation and temporal deduction to obtain the source-load power prediction value and key operation data deduction value of each node of the power grid in the future target period.

[0051] In some embodiments, the hybrid spatiotemporal graph neural network model is the core of the intelligence. This model is not a single algorithm, but a composite architecture integrating spatial understanding and temporal prediction capabilities. During model processing, it first analyzes the spatial relationships (topology awareness) between nodes along the time axis of the dynamic programming graph sequence at each "snapshot." Then, it fuses these spatiotemporal feature information to ultimately deduce the predicted values ​​of power generation output (source) and power demand (load) for each grid node at the future target time, as well as key operational status data such as voltage.

[0052] In some embodiments, the hybrid spatiotemporal graph neural network model includes a topology-aware module and a physical constraint inference module. The topology-aware module is constructed with a heterogeneous graph attention network as its core component. The input is a single topology in a dynamic programming graph sequence, containing node features (such as historical power and type encoding) and edge features (such as electrical parameters). Its core function is to perform spatial feature aggregation. Simulating the natural propagation of power flow in a power grid, it intelligently determines how much influence a node's state should be on its neighboring nodes by calculating the spatial attention weights between nodes. It considers the electrical parameters of the edges (such as impedance), enabling deeper information interaction between nodes with close electrical distances. Finally, this module generates a feature representation for each node that incorporates global topological correlations.

[0053] The physical constraint inference module uses a temporal convolutional network as its main architecture and embeds differentiable physical laws as regularization constraints. Its core function is to perform evolutionary predictions along the time dimension. It receives spatial features output from the topology-aware module and learns their change patterns along the time axis. During the learning process, this module introduces fundamental physical laws of the power grid (such as the power balance law) as constraints. By transforming the physical equations into differentiable loss function terms, it guides the model to learn physically feasible evolutionary paths, thereby ensuring that the predicted source-load power sequence is physically reliable and achievable, effectively avoiding absurd prediction results.

[0054] The two modules are connected in series to form an end-to-end learning framework that combines spatial understanding and temporal prediction. It can learn both spatial correlations and temporal trends from the dynamically changing power grid topology, and finally output the future source-load distribution and operating status that conforms to both data patterns and physical laws.

[0055] In some embodiments, the input layer of the hybrid spatiotemporal graph neural network model is constructed as follows: First, each graph structure (node ​​features, edge features, adjacency matrix) in the dynamic programming graph sequence is converted into a tensor format that the model can process. The topology-aware module is constructed by building a graph attention network layer. Differentiated feature mapping paths are designed based on the type of grid nodes (generators, loads, etc.). The network automatically learns how to calculate attention weights based on connectivity and edge parameters, and performs feature aggregation through multiple graph convolutional layers. The physical constraint deduction module is constructed by building a temporal convolutional network whose convolution kernels slide along the time axis to capture long-term dependencies. Simultaneously, differentiable grid physical equations (such as power balance equations) are defined at the code level and prepared as a module capable of calculating loss functions. The output layer is constructed using a multi-task learning mechanism, with parallel fully connected layers used for regressing and predicting active and reactive power, respectively. These layers can be connected to a lightweight linearized power flow calculation layer to simultaneously deduce voltage and phase angle.

[0056] In some embodiments, the training process of the hybrid spatiotemporal graph neural network model includes: Data preparation: Preparing high-quality power grid operation data from the same historical period as training samples, and dividing them into training set, validation set, and test set. Loss function definition: Defining the total loss function L as shown in the figure. total The function is a weighted sum of two parts: a prediction task loss (e.g., mean squared error, MSE), which measures the difference between the predicted and true values; and a physical constraint loss, which quantifies the degree to which the prediction violates physical laws. Model training and validation: The total loss function is minimized using a gradient descent algorithm (e.g., the Adam optimizer). In each training round, forward propagation calculates the predicted values ​​and the total loss, while backpropagation calculates the gradient and updates the model parameters. Training terminates early when performance on the validation set no longer improves to prevent overfitting, and the model's generalization performance is finally evaluated on the test set.

[0057] As one possible implementation, step S104 can be specifically implemented as steps S1041-S1046.

[0058] S1041. Based on the dynamic programming graph sequence, extract the connection relationship features between nodes and the electrical parameter features of edges in each dynamic programming graph.

[0059] For example, embodiments of the present invention can perform deep feature mining on the graph structure at each time step in a dynamic programming graph sequence. For node connectivity, not only are basic adjacency relationships extracted, but also topological importance indices such as degree centrality and betweenness centrality are analyzed to form structural feature vectors for the nodes. For edge electrical parameters, basic parameters such as line resistance, reactance, conductance, and susceptance are extracted, and derived features such as the natural power distribution coefficient and electrical distance of the line are further calculated. Simultaneously, considering the connectivity between devices at different voltage levels, a standardized feature representation across voltage levels is established, providing a comprehensive and standardized input feature set for subsequent neural network processing.

[0060] S1042. Based on the topology perception module of the hybrid spatiotemporal graph neural network model, and the connection relationship features between nodes and the electrical parameter features of edges in each dynamic programming graph, dynamically calculate the spatial attention weights between nodes in each dynamic programming graph.

[0061] For example, embodiments of the present invention can employ a multi-head attention mechanism to process different types of spatial correlations in parallel. Each attention head focuses on a specific type of inter-node relationship, such as electrical distance correlation, capacity matching correlation, and topological adjacency correlation. Through a trainable parameter matrix, node features and edge features are projected onto multiple subspaces, and correlation scores between nodes are calculated in each subspace. Finally, the calculation results of all attention heads are combined to generate an attention weight matrix reflecting multi-dimensional spatial dependencies. This weight matrix can dynamically adjust the strength of information transmission between nodes, accurately capturing the complex spatial correlation characteristics in the power grid.

[0062] For example, step S1042 can be specifically implemented as steps A1-A3.

[0063] A1. Based on the topology sensing module and the connection relationship characteristics between nodes, perceive the type differences of each node in each dynamic programming graph and calculate the basic correlation coefficient of each node.

[0064] For example, embodiments of the present invention can establish a node type compatibility matrix, which is predefined based on the physical connection rules and operating characteristics between power grid equipment. Differentiated basic association weights are set for connections between generation nodes and load nodes, and between nodes of different voltage levels. By analyzing the relative positions of nodes in the topology graph, their structural similarity is calculated, and combined with the operational coupling degree of node types, a basic association coefficient reflecting the inherent connection strength between nodes is comprehensively derived.

[0065] A2. Based on the electrical parameter characteristics of the edges, a trainable nonlinear mapping network is used to convert them into a modified shadow of the correlation between nodes.

[0066] For example, embodiments of the present invention can process the electrical parameter features of edges through a trainable nonlinear mapping network. This network employs a multilayer perceptron structure to nonlinearly transform continuous electrical parameters such as impedance, admittance, and rated capacity of the lines, outputting a correction factor characterizing the tightness of electrical connections. During training, the network automatically learns the complex mapping relationship between electrical parameters and node correlation, resulting in higher correlation enhancement for node pairs corresponding to lines with short electrical distances and strong transmission capabilities. This accurately reflects the actual impact of power grid physical characteristics on the spatial attention mechanism.

[0067] A3. Based on the basic correlation coefficient and correction factor of each node, perform weighted fusion to dynamically calculate the spatial attention weight of each node in each dynamic programming graph.

[0068] For example, in this embodiment of the invention, the basic correlation coefficient and the edge attribute correction factor can be multiplied element-wise to obtain the original attention score matrix. Then, an improved normalized exponential function is used to perform local normalization within the neighborhood of each node, ensuring that the sum of the attention weights of each node to all its neighboring nodes is 1. This local normalization method maintains the relative distribution characteristics of the attention weights while allowing the weight calculation to adapt to changes in the degree of different nodes, ultimately generating a spatial attention weight distribution that conforms to the characteristics of a graph structure.

[0069] S1043. Based on the spatial attention weights between nodes in each dynamic programming graph and the propagation strength of feature information between nodes, cross-topological aggregation of node features is achieved through multi-layer graph convolution operations to generate feature vectors that fuse global spatial relationships.

[0070] For example, embodiments of the present invention can implement a multi-level feature propagation mechanism based on attention weights. Through stacked multi-layer graph convolutional networks, each node sequentially aggregates feature information from its first-degree neighbors, second-degree neighbors, and so on, up to more distant nodes. During each convolutional layer, nodes selectively receive and weightedly fuse feature information from neighboring nodes according to spatial attention weights. As the network depth increases, each node can effectively capture its global position information within the entire topology and its potential associations with distant nodes, ultimately forming a node feature representation that includes both local details and a global perspective.

[0071] S1044. The physical constraint deduction module based on the feature vector of the fusion global spatial correlation and the hybrid spatiotemporal graph neural network model uses the differentiable physical equation as the regularization constraint to perform time evolution on the features of each node in the feature vector, so as to obtain the optimized time sequence features that satisfy the physical laws of the power grid.

[0072] For example, embodiments of the present invention can employ dilated causal convolutional networks to capture long-term temporal dependencies. During network training, in addition to optimizing prediction accuracy, regularization constraints based on the physical laws of power grid operation are introduced. These constraints are implemented through a differentiable computational graph and include power balance constraints, node voltage security constraints, and line transmission capacity constraints. By jointly optimizing the prediction loss and the physical constraint violation penalty term, the network is driven to spontaneously comply with the basic physical laws of power grid operation while learning temporal evolution patterns, ensuring the rationality and feasibility of the inference results in a physical sense.

[0073] For example, step S1044 can be specifically implemented as steps B1-B3.

[0074] B1. Based on the differentiable power balance constraint of Kirchhoff's law for power grids and the differentiable voltage safety constraint based on the power grid operation safety standard, determine the physical law regularization loss term.

[0075] For example, embodiments of the present invention can construct a power balance constraint loss term based on Kirchhoff's laws for power grids. This loss term calculates the algebraic sum of the injected power at each node, measuring its deviation from the theoretical zero value. Simultaneously, a voltage safety constraint loss term is constructed based on power grid operation safety standards. By setting upper and lower limits for voltage operation, a secondary penalty is imposed on voltage predictions that exceed these limits. These physical constraints are implemented through a differentiable computational graph, ensuring gradient backpropagation during training and transforming the hard physical constraints of power grid operation into soft optimization objectives.

[0076] B2. The physical law regularization loss term is weighted and fused with the loss function of the main prediction task of the physical constraint inference module to construct a multi-objective optimization loss function that integrates physical laws.

[0077] For example, embodiments of the present invention can employ an adaptive weighting strategy to balance the importance of different optimization objectives. The main prediction task loss function uses mean squared error to measure prediction accuracy, while the physical constraint loss term controls its influence through adjustable hyperparameters. In the early stages of training, the system focuses on optimizing prediction accuracy. As training progresses, the weight of physical constraints is gradually increased, guiding the model to gradually meet the physical requirements of power grid operation while ensuring prediction accuracy, thus forming a multi-objective optimization framework that considers both data-driven and physical principles.

[0078] B3. Based on the feature vector of global spatial correlation, guided by the optimized loss function, the gradient backpropagation algorithm is used to simultaneously optimize the prediction accuracy and the degree of conformity with physical laws, driving the physical constraint inference module to output optimized time series features that satisfy the physical laws of the power grid.

[0079] For example, embodiments of the present invention can simultaneously optimize multiple loss terms using a gradient backpropagation algorithm. During training, gradient information comes not only from the backpropagation of prediction errors but also from gradient signals indicating the degree of violation of physical constraints. This multi-source gradient jointly guides the update direction of model parameters, driving the physical constraint deduction module to learn a time-series evolution pattern that conforms to both historical data statistical patterns and power grid physical constraints, ultimately outputting physically reasonable optimized time-series characteristics.

[0080] S1045. Based on the optimized timing characteristics that satisfy the physical laws of the power grid, calculate the predicted source-load power of each node in parallel during the future target time period.

[0081] In some embodiments, the source-load power prediction values ​​include active power prediction values ​​and reactive power prediction values.

[0082] For example, embodiments of the present invention can employ a multi-task learning framework to simultaneously predict active and reactive power. The network output layer configures an independent dual-channel prediction head for each node, corresponding to the prediction tasks for active and reactive power respectively. The two prediction tasks share the spatiotemporal features extracted from the preceding layers but have their own specific output transformation parameters. This design ensures both collaborative prediction between active and reactive power and allows the network to learn the different variation characteristics of the two power types. Through end-to-end training, the model can fully utilize relevant information in the spatiotemporal features to achieve accurate parallel prediction of the power values ​​at each node.

[0083] S1046. Based on the predicted source-load power of each node in the future target period and the power grid topology parameters, key operational data projection values ​​are generated through linearized power flow calculation.

[0084] In some embodiments, the key operational data projection values ​​include projection values ​​of node voltage amplitude and phase angle.

[0085] For example, embodiments of the present invention can perform fast power flow calculations based on power prediction results. A linearized power flow calculation method is employed. First, a node admittance matrix is ​​formed based on grid topology parameters. Then, the node voltage phase angle is approximated using the active power prediction value through DC power flow. Based on this, the node voltage amplitude is estimated using a linearized voltage-power relationship equation, combining the reactive power prediction value and line parameters. The entire process employs a fully differentiable calculation flow, ensuring that the power flow calculation results can propagate back to influence front-end feature learning, forming a complete closed loop from power prediction to operational status calculation.

[0086] For example, step S1046 can be specifically implemented as steps C1-C4.

[0087] C1. For any dynamic programming graph and the target time in which the dynamic programming graph is located, construct a node admittance matrix for linearized power flow calculation based on the power grid topology and line parameters of the dynamic programming graph at the target time.

[0088] For example, embodiments of the present invention can analyze the connection relationships and parameter information of all lines and transformers based on the power grid topology at a target time. The admittance values ​​of the lines are calculated according to their impedance parameters and filled into the corresponding positions of the admittance matrix according to the node connection relationships. For transformer branches, their turns ratio parameters are standardized to form a complete node admittance matrix.

[0089] C2. Based on the node admittance matrix and the predicted active power of each node at the target time, the DC power flow method is used to solve for the derived voltage phase angle of each node.

[0090] For example, embodiments of the present invention can employ a simplified calculation principle of the DC power flow method. Ignoring the effects of line resistance and ground admittance, the AC power flow equations are linearized, establishing a linear relationship between node active power injection and voltage phase angle. By solving the system of linear equations composed of the imaginary parts of the admittance matrix, the derived values ​​of the voltage phase angle at each node are directly calculated. This method significantly improves computational efficiency while maintaining computational accuracy, making it suitable for rapid analysis and calculation of large-scale power grids.

[0091] C3. Based on the derived voltage phase angle of each node and the predicted reactive power of each node at the target time, the derived voltage amplitude of each node is estimated through the linear voltage amplitude equation.

[0092] For example, embodiments of the present invention can estimate the voltage amplitude based on the obtained voltage phase angle derivation value, combined with line parameters and node reactive power prediction values, through a linearized voltage amplitude equation. This method considers the main influence of reactive power flow on voltage amplitude, while also taking into account the coupling effect of the voltage phase angle difference between adjacent nodes on the local voltage. It quickly estimates the voltage amplitude derivation value of each node through a linear approximation relationship, forming a complete power grid operating state estimate.

[0093] C4. Based on the voltage phase angle and voltage amplitude of each node corresponding to each dynamic programming diagram, generate the key operational data projection values.

[0094] For example, in embodiments of the present invention, the calculated voltage phase angle and voltage amplitude projection values ​​can be integrated according to node numbers and stored in association with the corresponding topology and timestamp information. For each target time corresponding to a dynamic programming diagram, a data set containing the complete operating status of all nodes in the entire network is generated. This data constitutes the main content of the key operating data projection values, providing complete operating status information for subsequent power grid planning analysis and theoretical line loss calculation.

[0095] This invention provides a power grid source-load prediction method based on a spatiotemporal graph neural network. By parsing medium- and long-term planning schemes into graph evolution operation instructions and generating a dynamic programming graph sequence, a complete mathematical description of the power grid growth process is achieved. This allows the prediction model to accurately match the temporal path of power grid planning and construction, solving the problem of mismatch between static models and dynamic planning. Subsequently, based on the dynamic programming graph sequence, combined with a hybrid spatiotemporal graph neural network, spatial feature aggregation and temporal extrapolation are performed. Considering the spatial constraints brought about by power grid topology changes and the long-term evolution law of source and load, high-quality and high-reliability source-load prediction data is output. This solves the problem of disconnect between static power grid topology and reality in source-load prediction, improving the accuracy of power grid source-load prediction.

[0096] Optionally, the power grid source-load prediction method based on spatiotemporal graph neural network provided in this embodiment of the invention further includes steps S201-S205.

[0097] S201. Obtain the approval status, funding guarantee level, and expected construction period of each planning project in the medium- and long-term planning scheme.

[0098] S202. Based on the approval status, funding guarantee level, and expected construction period of each planning project, use a weighted evaluation model to determine the confidence level of the successful implementation of each planning project.

[0099] Exemplarily, embodiments of the present invention can adopt a multi-dimensional weighted evaluation system to quantitatively analyze the implementation feasibility of each planning project. For the approval status dimension, set corresponding stage completion coefficients according to the approval stage of the project (such as project establishment, feasibility study, approval, under construction, etc.); for the funding guarantee dimension, comprehensively consider the reliability of the funding source, the proportion of funds in place, and the clarity of the investment plan; for the construction period dimension, analyze the rationality of the project construction period arrangement, the complexity of the critical path, and the progress of the preliminary preparation work. By establishing an evaluation model with adjustable weights, the qualitative evaluations of these three dimensions are transformed into a unified confidence quantification index, forming a scientific prediction of the implementation possibility of the planning project.

[0100] S203. Based on the confidence level of the successful implementation of each planning project, sample and perturb the dynamic planning graph sequence to generate multiple differential planning graph sequences corresponding to the baseline scenario, optimistic scenario, and conservative scenario, as well as the implementation confidence weight of each scenario.

[0101] Exemplarily, embodiments of the present invention can implement a differential sequence generation strategy based on the confidence evaluation result. For the baseline scenario, a deterministic sequence that completely follows the planning timeline is generated; for the optimistic scenario, the commissioning time of high-confidence projects is advanced based on the baseline scenario, and the construction progress of related projects is increased; for the conservative scenario, the implementation time of low-confidence projects is postponed, and the phased construction or scale reduction of some projects is considered. The generation of each scenario maintains the integrity of the topological structure and logical consistency, ensuring that the generated differential planning graph sequences not only reflect different development possibilities but also conform to the basic laws of power grid construction.

[0102] S204. Input each differential planning graph sequence into the hybrid spatio-temporal graph neural network model for parallel deduction to obtain the deduction result.

[0103] In some embodiments, the deduction result includes the source-load power prediction value and the key operation data deduction value under each scenario.

[0104] For example, embodiments of the present invention can utilize a distributed computing architecture to simultaneously process prediction tasks for multiple scenarios. Each scenario's planning graph sequence is allocated independent computing resources while sharing the parameter weights of the same hybrid spatiotemporal graph neural network model. During the derivation process, the system records intermediate results and feature representations for each scenario, providing detailed process data for subsequent result comparison and analysis. This parallel processing mechanism not only improves computational efficiency but, more importantly, ensures the comparability of derivation results for different scenarios, avoiding systemic biases caused by sequential computation.

[0105] S205. Based on the implementation confidence weights of each scenario, the inference results are weighted and fused to generate source-load prediction intervals and key operational data confidence ranges with probability distribution characteristics.

[0106] For example, embodiments of the present invention may employ a result fusion algorithm based on confidence weights. First, the inference results for each scenario are standardized to eliminate the influence of different units; then, based on the confidence weights of each scenario, the predicted values ​​of the same indicator are weighted and averaged to obtain the expected value estimate; further, the distribution characteristics of the results for each scenario are analyzed, and their statistical variance and confidence interval are calculated.

[0107] Thus, by introducing a confidence assessment mechanism for planning projects and a multi-scenario parallel extrapolation mechanism, this invention elevates traditional deterministic prediction to a probabilistic prediction level. This invention quantifies the uncertainty of planning implementation, generates prediction results with confidence intervals, enabling planners to accurately grasp the source-load distribution range under different development scenarios. This provides risk-aware decision support for power grid planning, significantly improving the adaptability of planning schemes to future uncertain environments and the scientific rigor of decision-making.

[0108] Optionally, the power grid source-load prediction method based on spatiotemporal graph neural network provided in this embodiment of the invention further includes steps S301-S304.

[0109] S301. Based on the dynamic programming graph sequence and the randomly generated extreme disturbance scenario, generate an extreme programming graph sequence.

[0110] In some embodiments, the extreme disturbance scenarios include extreme weather events and equipment failure scenarios.

[0111] For example, embodiments of the present invention can establish a multi-probability level extreme event scenario library based on historical disaster statistics and equipment operational reliability records. For extreme weather events, considering historical observation data from meteorological departments and climate prediction models, weather scenarios of different intensities such as typhoons, hail, and rainstorms are generated, and corresponding spatial impact range and duration parameters are configured for each scenario. For equipment failure scenarios, based on the equipment's service life, maintenance records, and family defect information, multiple failure combination scenarios, including N-1 and N-2, are constructed, while also considering complex situations such as correct operation and maloperation or failure to operate of relay protection. Each scenario is assigned an estimated probability of occurrence and a severity rating, forming a complete knowledge base of extreme disturbance scenarios.

[0112] For example, embodiments of the present invention can employ a scene-injection-based dynamic graph evolution method. The selected extreme disturbance scenario is spatiotemporally aligned with the original dynamic programming graph sequence to simulate the impact of the disturbance event on the power grid topology at the target time step. Based on the spatial distribution characteristics of extreme weather, the fault probability of affected lines is calculated, and the set of lines actually interrupted is determined based on Monte Carlo sampling. For equipment fault scenarios, the corresponding nodes and edges are directly removed to simulate the state of equipment out of operation. Simultaneously considering the action logic of protection devices and the response of automatic power grid safety devices, a disturbed programming graph sequence reflecting the actual state of the power grid after the disturbance is generated.

[0113] S302. Based on the extreme planning graph sequence and the hybrid spatiotemporal graph neural network model, extrapolation is performed to obtain the extrapolation results under extreme perturbation scenarios.

[0114] For example, embodiments of the present invention can input the disturbed planning graph sequence into a hybrid spatiotemporal graph neural network model for specialized training and inference analysis. The model fully learns the operating characteristics of the power grid under extreme conditions, including emergency response mechanisms such as load transfer paths, backup power supply activation, and network reconfiguration strategies. During the inference process, the model not only considers normal power balance constraints but also introduces special constraints under emergency operating conditions, such as line overload tolerance and emergency voltage control strategies, to ensure that the inference results truly reflect the actual response capability and operating status of the power grid under extreme disturbances.

[0115] S303. Based on the simulation results under the extreme disturbance scenario, determine the resilience index of each node of the power grid in the future target period.

[0116] In some embodiments, the resilience metrics include power restoration time, load loss range, and the proportion of renewable energy disconnection from the grid.

[0117] For example, embodiments of the present invention can establish a multi-dimensional resilience assessment index system. The power restoration time index calculates the average time required from the occurrence of a disturbance to the restoration of 95% of the load by analyzing the timing of load restoration in the simulation results. The load loss range index counts the proportion of the total load that cannot be restored during the disturbance period to the total load of the entire network. The renewable energy disconnection ratio index calculates the proportion of renewable energy generator capacity disconnected due to frequency fluctuations, voltage exceeding limits, etc., to the total installed capacity of renewable energy. Simultaneously, auxiliary indicators such as critical load guarantee rate and network connectivity are considered to comprehensively quantify the resilience level of the power grid.

[0118] S304. Based on the resilience indicators of each node of the power grid in the future target period, identify the weak links in the power grid planning and determine the weak projects and weak nodes in the medium and long-term planning scheme.

[0119] For example, embodiments of the present invention can employ a multi-scenario cross-analysis identification method. By comparing and analyzing resilience indicators under different extreme disturbance scenarios, grid components and regions that perform poorly in most scenarios are identified. Through topology sensitivity analysis, critical lines and nodes whose failures would lead to large-scale load losses are determined; through electrical parameter analysis, weak links with insufficient voltage support and prominent transmission bottlenecks are identified. Finally, a complete assessment report is generated, including a list of weak items, a distribution map of weak nodes, and improvement priorities, providing precise decision support for optimizing and improving planning schemes.

[0120] Thus, this invention achieves a forward-looking assessment of the disaster resistance capability of the planned power grid by constructing extreme disturbance scenarios and performing resilience simulations. This invention extends planning analysis from normal operating conditions to extreme fault conditions, enabling precise identification of weaknesses in planning schemes, quantitative assessment of the power grid's resilience level, and providing specific improvement directions for enhancing the power grid's disaster prevention and mitigation capabilities, effectively improving the safety, resilience, and reliability of power grid planning.

[0121] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0122] The following are device embodiments of the present invention. For details not described in detail, please refer to the corresponding method embodiments described above.

[0123] Figure 2 A schematic diagram of a power grid source-load prediction device based on a spatiotemporal graph neural network according to an embodiment of the present invention is shown. The prediction device 400 includes a communication module 401 and a processing module 402.

[0124] The communication module 401 is used to acquire medium- and long-term planning schemes for the power grid and current power grid topology data.

[0125] The processing module 402 is used to perform power grid structure change analysis based on the medium- and long-term planning scheme to obtain multiple graph evolution operation instructions; based on the multiple graph evolution operation instructions and the current power grid topology data, it performs serialization deduction to generate a dynamic planning graph sequence of the power grid topology from the current basic state to the future planning state, the dynamic planning graph sequence being used to characterize the dynamic evolution process of the power grid topology over time; based on the dynamic planning graph sequence and a pre-built hybrid spatiotemporal graph neural network model, it performs spatial feature aggregation and temporal deduction to obtain the predicted source-load power values ​​and key operational data deduction values ​​of each node in the power grid within the future target time period.

[0126] Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. The electronic device 500 includes: a processor 501, a memory 502, and a computer program 503 stored in the memory 502 and executable on the processor 501. When the processor 501 executes the computer program 503, it implements the steps in the above-described method embodiments. Alternatively, when the processor 501 executes the computer program 503, it implements the functions of each module / unit in the above-described device embodiments.

[0127] For example, the computer program 503 may be divided into one or more modules / units, which are stored in the memory 502 and executed by the processor 501 to complete the present invention. The one or more modules / units may be a series of computer program instruction segments capable of performing specific functions, which describe the execution process of the computer program 503 in the electronic device 500.

[0128] The processor 501 may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor.

[0129] The memory 502 can be an internal storage unit of the electronic device 500, such as a hard disk or memory of the electronic device 500. The memory 502 can also be an external storage device of the electronic device 500, such as a plug-in hard disk, smart media card (SMC), secure digital card (SD), flash card, etc., equipped on the electronic device 500. Furthermore, the memory 502 can include both internal and external storage units of the electronic device 500. The memory 502 is used to store the computer program and other programs and data required by the terminal. The memory 502 can also be used to temporarily store data that has been output or will be output.

[0130] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A power grid source-load prediction method based on a spatio-temporal graph neural network, characterized in that, include: Obtain medium- and long-term power grid planning schemes and current power grid topology data; Based on the aforementioned medium- and long-term planning scheme, the power grid structure change analysis is performed to obtain multiple graph evolution operation instructions; Based on the multiple graph evolution operation instructions and the current power grid topology data, a serialized deduction is performed to generate a dynamic planning graph sequence of the power grid topology from the current basic state to the future planned state. The dynamic planning graph sequence is used to characterize the dynamic evolution process of the power grid topology over time. Based on the dynamic programming graph sequence and the pre-built hybrid spatiotemporal graph neural network model, spatial feature aggregation and temporal deduction are performed to obtain the source-load power prediction values ​​and key operation data deduction values ​​of each node of the power grid within the future target time period.

2. The power grid source-load prediction method based on the spatio-temporal graph neural network according to claim 1, characterized in that, Based on the aforementioned medium- and long-term planning scheme, the power grid structure change analysis is performed to obtain multiple graph evolution operation instructions, including: Based on the text and drawings of the aforementioned medium- and long-term planning scheme, key planning elements for power grid structure changes are identified and extracted; these key planning elements include power generation equipment, substation equipment, load nodes to be added or decommissioned, and transmission lines to be newly built or renovated. Each key planning element is analyzed, and the graph evolution operation instruction type is mapped to a graph evolution atomic operation instruction. The graph evolution atomic operation instruction includes adding a node, deleting a node, adding an edge, deleting an edge, modifying node attributes, and modifying edge attributes. Execution parameters are set for each graph evolution atomic operation instruction to obtain the operation instruction after setting the parameters. Among them, the execution parameters of the node-related instructions in the operation instruction include the node's unique identifier, node type, and planned capacity; the execution parameters of the edge-related instructions in the operation instruction include the identifier pair of the source node and the target node, the line type, and the planned impedance. Based on the project commissioning time sequence and the logical dependency relationship between the grid connection in the medium and long-term planning scheme, the operation instructions after setting parameters are sorted by time sequence and logically grouped to generate multiple graph evolution operation instructions with a clear execution order.

3. The power grid source-load prediction method based on spatiotemporal graph neural network according to claim 1, characterized in that, The step of performing serialized deduction based on the multiple graph evolution operation instructions and the current power grid topology data to generate a dynamic planning graph sequence of the power grid topology from the current basic state to the future planned state includes: Based on the current power grid topology data, an initial state diagram is constructed as the starting point for the deduction; Based on the aforementioned medium- and long-term planning scheme, a target planning map is constructed as the focus of the simulation. The multiple graph evolution operation instructions are sorted according to the project commissioning time, and operation instructions belonging to the same timestamp are grouped into the same instruction group to obtain a time sequence instruction group sequence. Starting from the initial state diagram, the timing instruction group sequence is executed sequentially in chronological order to obtain multiple intermediate state diagrams. After each instruction group is successfully executed, the power grid topology diagram is updated once, and the updated power grid topology diagram is used as the intermediate state diagram after the execution of that instruction group. Based on the initial state diagram, the multiple intermediate state diagrams, and the target planning diagram, they are combined in chronological order to form a dynamic planning diagram sequence.

4. The power grid source-load prediction method based on spatiotemporal graph neural network according to claim 1, characterized in that, Based on the dynamic programming graph sequence and the pre-constructed hybrid spatiotemporal graph neural network model, spatial feature aggregation and temporal extrapolation are performed to obtain the predicted source-load power values ​​and key operational data extrapolation values ​​of each node in the power grid within the future target time period, including: Based on the dynamic programming graph sequence, extract the connection relationship features between nodes and the electrical parameter features of edges in each dynamic programming graph; Based on the topology perception module of the hybrid spatiotemporal graph neural network model, as well as the connection relationship features between nodes and the electrical parameter features of edges in each dynamic programming graph, the spatial attention weights between nodes in each dynamic programming graph are dynamically calculated. Based on the spatial attention weights between nodes in each dynamic programming graph and the propagation strength of feature information between nodes, cross-topological aggregation of node features is achieved through multi-layer graph convolution operations to generate feature vectors that fuse global spatial relationships. Based on the feature vectors that integrate global spatial correlation and the physical constraint deduction module of the hybrid spatiotemporal graph neural network model, the time evolution of the features of each node in the feature vector is performed using differentiable physical equations as regularization constraints to obtain optimized time-series features that satisfy the physical laws of the power grid. Based on the optimized timing characteristics that satisfy the physical laws of the power grid, the source-load power prediction values ​​of each node in the future target time period are calculated in parallel. The source-load power prediction values ​​include active power prediction values ​​and reactive power prediction values. Based on the predicted source-load power of each node in the future target period and the grid topology parameters, key operational data projection values ​​are generated through linearized power flow calculation. These key operational data projection values ​​include the projection values ​​of node voltage amplitude and phase angle.

5. The power grid source-load prediction method based on spatiotemporal graph neural network according to claim 4, characterized in that, The topology-aware module based on the hybrid spatiotemporal graph neural network model, along with the connection relationship features between nodes and the electrical parameter features of edges in each dynamic programming graph, dynamically calculates the spatial attention weights between nodes in each dynamic programming graph, including: Based on the topology sensing module and the characteristics of the connection relationship between nodes, the type differences of each node in each dynamic programming graph are sensed, and the basic correlation coefficient of each node is calculated. Based on the electrical parameter characteristics of the edges, a trainable nonlinear mapping network is used to convert them into a corrected shadow of the correlation between nodes. Based on the basic correlation coefficient and correction factor of each node, a weighted fusion is performed to dynamically calculate the spatial attention weight of each node in each dynamic programming graph.

6. The power grid source-load prediction method based on spatiotemporal graph neural network according to claim 4, characterized in that, The physical constraint deduction module based on the feature vector fused with global spatial correlation and the hybrid spatiotemporal graph neural network model uses differentiable physical equations as regularization constraints to perform time evolution on the features of each node in the feature vector, obtaining optimized time-series features that satisfy the physical laws of the power grid, including: Based on the differentiable power balance constraint of the power grid Kirchhoff's law and the differentiable voltage safety constraint based on the power grid operation safety standard, the physical law regularization loss term is determined. The physical law regularization loss term is weighted and fused with the loss function of the main prediction task of the physical constraint inference module to construct a multi-objective optimization loss function that integrates physical laws. Based on the feature vectors of the fusion global spatial correlation, guided by the optimized loss function, the gradient backpropagation algorithm simultaneously optimizes the prediction accuracy and the degree of conformity with physical laws, driving the physical constraint inference module to output optimized time-series features that satisfy the physical laws of the power grid.

7. The power grid source-load prediction method based on spatiotemporal graph neural network according to claim 4, characterized in that, Based on the predicted source-load power of each node in the future target time period and the power grid topology parameters, key operational data projection values ​​are generated through linearized power flow calculations, including: For any dynamic programming graph and the target time in which the dynamic programming graph is located, a node admittance matrix for linearized power flow calculation is constructed based on the power grid topology and line parameters of the dynamic programming graph at the target time. Based on the node admittance matrix and the predicted active power of each node at the target time, the DC power flow method is used to solve for the derived voltage phase angle of each node. Based on the derived voltage phase angle of each node and the predicted reactive power of each node at the target time, the derived voltage amplitude of each node is estimated through the linear voltage amplitude equation. Based on the voltage phase angle and voltage amplitude of each node corresponding to each dynamic programming diagram, the key operational data projection values ​​are generated.

8. The power grid source-load prediction method based on spatiotemporal graph neural network according to claim 1, characterized in that, The method further includes: Obtain the approval status, funding guarantee level, and expected construction period of each planning project in the medium- and long-term planning scheme; Based on the approval status, funding guarantee level, and expected construction period of each planning project, a weighted evaluation model is used to determine the confidence level of the successful implementation of each planning project. Based on the confidence level of the successful implementation of each planning project, the dynamic planning graph sequence is sampled and perturbed to generate multiple differentiated planning graph sequences corresponding to the baseline scenario, optimistic scenario and conservative scenario, respectively, as well as the implementation confidence weight of each scenario. Each differentiated planning map sequence is input into the hybrid spatiotemporal graph neural network model for parallel extrapolation to obtain extrapolation results, which include source-load power prediction values ​​and key operational data extrapolation values ​​under each scenario; Based on the implementation confidence weights for each scenario, the inference results are weighted and fused to generate source-load prediction intervals and key operational data confidence ranges with probability distribution characteristics.

9. The power grid source-load prediction method based on spatiotemporal graph neural network according to claim 1, characterized in that, The method further includes: Based on the dynamic programming graph sequence and randomly generated extreme disturbance scenarios, an extreme programming graph sequence is generated, wherein the extreme disturbance scenarios include extreme weather events and equipment failure scenarios; Based on the extreme programming graph sequence and the hybrid spatiotemporal graph neural network model, the deduction results under extreme perturbation scenarios are obtained. Based on the simulation results under the extreme disturbance scenario, the resilience indicators of each node of the power grid in the future target period are determined. The resilience indicators include power supply restoration time, load loss range and the proportion of new energy disconnection from the grid. Based on the resilience indicators of each node of the power grid in the future target period, weak links in the power grid planning are identified, and weak projects and weak nodes in the medium and long-term planning scheme are determined.

10. A power grid source-load prediction device based on a spatiotemporal graph neural network, characterized in that, include: The communication module is used to acquire medium- and long-term power grid planning schemes and current power grid topology data; The processing module is used to perform power grid structure change analysis based on the medium- and long-term planning scheme and obtain multiple graph evolution operation instructions; Based on the multiple graph evolution operation instructions and the current power grid topology data, a serialized deduction is performed to generate a dynamic planning graph sequence of the power grid topology from the current basic state to the future planned state. The dynamic planning graph sequence is used to characterize the dynamic evolution process of the power grid topology over time. Based on the dynamic planning graph sequence and a pre-built hybrid spatiotemporal graph neural network model, spatial feature aggregation and temporal deduction are performed to obtain the predicted source-load power values ​​and key operational data deduction values ​​of each node in the power grid within the future target time period.

Images