An intelligent power grid load prediction method and system based on graph computing

Abstract

The application provides a kind of smart grid load forecasting method and system of graph computing, it is related to load forecasting technical field, including: constructing power grid load topology;Degree connection relationship analysis is carried out to N load nodes, and N load nodes are divided into K load demand nodes and M load distribution nodes;Pre-construct K load single-point prediction model;And construct M load correlation prediction model;K load single-point prediction model and M load correlation prediction model complete the construction of load multilevel graph computing network;Through real-time acquisition K update single-node time series data, load prediction is cyclically updated, and power supply source load output is obtained.Through the application, the technical problem that the load prediction result is inaccurate due to the neglect of the transmission relationship between different node loads in the power grid in the prior art can be solved, and the accuracy of load prediction is improved by constructing a load multilevel graph computing network.

Description

Technical Field

[0001] The present application relates to the technical field of load forecasting, and in particular to a graph-based smart grid load forecasting method and system. Background Art

[0002] Load forecasting is a core component of smart grid operations, directly impacting grid dispatch strategies, power resource allocation, and the stability and economic efficiency of equipment operation. Smart grid load forecasting encompasses short-term load forecasting (e.g., hourly and daily forecasts), medium-term load forecasting (e.g., weekly and monthly forecasts), and long-term load forecasting (e.g., annual forecasts). Currently, most existing load forecasting methods build models based on the temporal correlation of historical load series. However, these methods often overlook the spatial coupling between adjacent nodes in the grid and the combined impact of external factors on load fluctuations. Loads at different nodes exhibit strong spatiotemporal correlations, such as those caused by dynamic correlations due to transmission line connections or load transfer patterns. Failure to fully account for these spatiotemporal couplings will directly impact the accuracy of load forecasting.

[0003] In summary, the prior art has a technical problem in that the load forecasting result is inaccurate due to ignoring the transfer relationship between loads at different nodes in the power grid. Summary of the Invention

[0004] The purpose of this application is to provide a graph-based smart grid load forecasting method and system to solve the technical problem in the prior art of inaccurate load forecasting results due to ignoring the transfer relationship between loads at different nodes in the power grid.

[0005] In view of the above problems, the present application provides a smart grid load forecasting method and system based on graph computing.

[0006] In the first aspect, the present application provides a graph-based smart grid load forecasting method, which is implemented by a graph-based smart grid load forecasting system, wherein the graph-based smart grid load forecasting method includes: constructing a grid load topology by performing load analysis on a target grid; performing in-degree connection relationship analysis on N load nodes in the grid load topology, dividing the N load nodes into K load demand nodes and M load distribution nodes, wherein K+M=N, K>M; pre-constructing K load single-point forecasting models for the K load demand nodes; calling the M groups of out-degree topological relationships of M load distribution nodes are obtained, and M load association prediction models of the M load distribution nodes are constructed based on the M groups of out-degree topological relationships; the K load single-point prediction models are mapped and loaded to the K load demand nodes of the power grid load topology, and the M load association prediction models are mapped and loaded to the M load distribution nodes of the power grid load topology to complete the construction of the load multi-level graph calculation network; K updated single-node time series data of the K load demand nodes are collected in real time, and the K updated single-node time series data are synchronized to the load multi-level graph calculation network for load forecasting cycle update to obtain the power supply source load output.

[0007] In the second aspect, the present application also provides a graph-computing smart grid load forecasting system for executing a graph-computing smart grid load forecasting method as described in the first aspect, wherein the graph-computing smart grid load forecasting system includes: a load analysis module, the load analysis module is used to construct a grid load topology by performing load analysis on the target grid; a connection relationship analysis module, the connection relationship analysis module is used to perform in-degree connection relationship analysis on N load nodes in the grid load topology, and divide the N load nodes into K load demand nodes and M load distribution nodes, wherein K+M=N, K>M; a first model construction module, the first model construction module is used to pre-construct K load single-point prediction models of the K load demand nodes; a second model construction module, the second model construction module Used to call the M groups of out-degree topological relationships of the M load distribution nodes in the power grid load topology, and construct M load association prediction models of the M load distribution nodes based on the M groups of out-degree topological relationships; a graph computing network construction module, the graph computing network construction module is used to map and load the K load single-point prediction models to the K load demand nodes of the power grid load topology, and map and load the M load association prediction models to the M load distribution nodes of the power grid load topology, to complete the construction of the load multi-level graph computing network; a load prediction module, the load prediction module is used to collect K updated single-node time series data of the K load demand nodes in real time, and synchronize the K updated single-node time series data to the load multi-level graph computing network for load prediction cycle update to obtain the power supply load output.

[0008] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0009] By performing load analysis on the target power grid, a power grid load topology is constructed; an in-degree connection relationship analysis is performed on N load nodes in the power grid load topology, and the N load nodes are divided into K load demand nodes and M load distribution nodes, where K+M=N, K>M; K load single-point prediction models of the K load demand nodes are pre-constructed; M groups of out-degree topological relationships of the M load distribution nodes are called in the power grid load topology, and M load association prediction models of the M load distribution nodes are constructed according to the M groups of out-degree topological relationships; the K load single-point prediction models are mapped and loaded to the K load demand nodes of the power grid load topology, and the M load association prediction models are mapped and loaded to the M load distribution nodes of the power grid load topology, thereby completing the construction of a load multi-level graph computing network; K updated single-node time series data of the K load demand nodes are collected in real time, and the K updated single-node time series data are synchronized to the load multi-level graph computing network for load prediction cycle update, thereby obtaining the power supply load output. That is to say, by constructing the power grid load topology, the load transfer relationship between nodes is converted into the in-degree and out-degree relationship in the graph structure. For the demand node, a single-point load prediction model is constructed to capture local load changes; for the distribution node, its out-degree relationship is used to construct a load association prediction model to accurately characterize the load transfer law between nodes, avoid the deviation caused by the independence assumption of the single-point prediction model, and improve the accuracy of load prediction.

[0010] The above description is only an overview of the technical solution of the present application. In order to more clearly understand the technical means of the present application, which can be implemented in accordance with the contents of the description, and to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the specific implementation methods of the present application are specifically listed below. It should be understood that the content described in this section is not intended to identify the key or important features of the embodiments of the present application, nor is it intended to limit the scope of the present application. Other features of the present application will become easy to understand through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] In order to more clearly illustrate the technical solutions in this application or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are merely exemplary, and a person of ordinary skill in the art can obtain other drawings based on the provided drawings without any creative work.

[0012] Figure 1 This is a flow chart of a graph computing method for smart grid load forecasting in this application;

[0013] Figure 2 This is a structural diagram of a graph computing smart grid load forecasting system for this application.

[0014] Explanation of the accompanying symbols: load analysis module 11, connection relationship analysis module 12, first model construction module 13, second model construction module 14, graph computing network construction module 15, load forecasting module 16. DETAILED DESCRIPTION

[0015] This application provides a graph-based smart grid load forecasting method and system, addressing the existing technical problem of inaccurate load forecasting results due to ignoring the transfer relationship between loads at different nodes in the power grid. By constructing a power grid load topology, the load transfer relationship between nodes is converted into the in-degree and out-degree relationships in the graph structure. For demand nodes, a single-point load forecasting model is constructed to capture local load changes. For distribution nodes, a load association forecasting model is constructed using their out-degree relationships to accurately depict the load transfer patterns between nodes, avoiding the deviation caused by the independence assumption of the single-point forecasting model and improving the accuracy of load forecasting.

[0016] Below, the technical solutions in this application will be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of this application, rather than all of the embodiments of this application. It should be understood that this application is not limited to the example embodiments described herein. Based on the embodiments of this application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application. It should also be noted that, for the sake of ease of description, only the parts related to this application, rather than all of them, are shown in the accompanying drawings.

[0017] For example 1, please refer to the attached Figure 1 The present application provides a graph-based smart grid load forecasting method, wherein the graph-based smart grid load forecasting method is applied to a graph-based smart grid load forecasting system, and the graph-based smart grid load forecasting method specifically includes the following steps:

[0018] S100: Analyze the load of the target power grid and construct a power grid load topology.

[0019] Specifically, a load analysis is performed on the target power grid, collecting historical load data from each node and analyzing the load trends, fluctuations, and other characteristics of each node. H nodes in the target power grid are extracted, and historical load data and historical operation and maintenance data corresponding to each node are obtained. Load fluctuation analysis and fault impact analysis are performed separately to determine the load fluctuation frequency, load fluctuation scale, fault radiation coverage, and fault interval scale. The power grid is traversed, and the hierarchical node connectivity of each node is calculated. Power flow calculations are performed on the H nodes to assess the influence of each node. The specific calculation process is explained in detail in the corresponding weighted steps below. Here, we will only introduce the concept and not elaborate on the details. The load fluctuation frequency reflects the temporal variation of the node load; the load fluctuation scale indicates the magnitude of the load fluctuation; the fault radiation coverage refers to the extent of the impact of a fault on other nodes; the fault interval scale is the time interval between two faults; the node centrality reflects the importance of the node in the network; and the node influence indicates the impact of the node on the overall power grid performance.

[0020] By analyzing the load fluctuation frequency, load fluctuation scale, fault radiation coverage, fault interval scale, node centrality, and node influence of each node, the contribution of each node is calculated. Based on the contribution of each node, the N load nodes with the greatest influence are selected. The power flow path analysis is performed on the N load nodes to determine the connection method and power flow path in the power grid. Based on the results of the power flow path analysis, the N load nodes are connected to form a power grid load topology graph. A topology graph is a directed graph in which nodes represent load nodes in the power grid and edges represent the direction of power flow. For example, if node A points to an edge at node B, it means that the power of node A flows to node B. The load change at node A will affect node B, but the load change at node B will not directly affect node A.

[0021] By leveraging the connections within the directed graph within the grid load topology, the load demands of other nodes can be inferred based on the load demands of known nodes. For example, if the load demand of node B is known, the load demand of node A can be inferred based on the topological relationships. By analyzing the grid load topology and its connections, load changes can be predicted, allowing for more precise scheduling based on the load demands of each node and the power flow path, thereby improving grid operation efficiency.

[0022] S200: Performing in-degree connection relationship analysis on N load nodes in the power grid load topology, dividing the N load nodes into K load demand nodes and M load distribution nodes, where K+M=N, K>M.

[0023] Specifically, the in-degree connectivity of N load nodes in the power grid load topology is analyzed. The in-degree of each load node in the power grid load topology is analyzed. This means the number of arrows pointing to that node represents how many other nodes are supplying power to it. In a directed graph, in-degree represents the number of edges pointing to that node. In graph theory, the in-degree of a node refers to the number of edges pointing to that node. In-degree connectivity analysis analyzes the direction of connections between each load node and other nodes in the power grid topology, determining how much power from other nodes flows to each load node.

[0024] Based on the results of in-degree connectivity analysis, N load nodes are divided into K load demand nodes and M load distribution nodes. Based on the role of each node in the power grid, load demand nodes are those that directly draw power from the grid, while load distribution nodes are responsible for distributing power to different regions or load demand nodes. K represents the number of load demand nodes, and M represents the number of load distribution nodes, with K + M = N. Typically, the number of load demand nodes, K, is greater than the number of load distribution nodes, M, because the number of loads in a power grid is typically greater than that of substations or distribution stations. The majority of nodes in a power grid are load demand nodes, with only a few (such as substations) performing load distribution functions. Through in-degree connectivity analysis, we can identify nodes that play a critical role in load demand in the power grid. Load changes at load demand nodes warrant prioritization, while the scheduling of load distribution nodes can be adjusted based on the distribution situation, thereby ensuring grid operating efficiency and load balance.

[0025] S300: Pre-constructing K load single-point prediction models for the K load demand nodes.

[0026] Specifically, multiple historical load records and multiple historical external data (such as weather, time, and holidays) are obtained for each load demand node. Taking the first load demand node as an example, the sliding time window technique is used to process the multiple historical external data for the first load demand node, dividing them into multiple sample single-node time series data. Each sample single-node time series data is identified by a last timestamp. Based on the last timestamp, the sample instantaneous load demand (i.e., the load demand at the last moment) is retrieved from the multiple historical load records to obtain multiple sample instantaneous load demands.

[0027] Data preprocessing is performed on multiple sample single-node time series data and multiple sample instantaneous load demands corresponding to the first load demand node. A prediction model suitable for processing time series data, such as an LSTM model, is selected. The network structure, including the number of layers, the number of neurons in each layer, and the activation function, is determined to construct a first load single-point prediction model. The preprocessed data is input into the first load single-point prediction model for training. The model will learn how to predict load demand based on the input time series data. During training, a validation set is used to evaluate the model's performance. If the model's performance does not meet expectations, parameters such as the network structure and learning rate can be adjusted and training can continue. The above training process is repeated to train load single-point prediction models for the second load demand node through the Kth load demand node, ultimately obtaining K load single-point prediction models. By constructing a prediction model based on the historical data and external data of each load demand node, the grid load demand can be effectively predicted, the dependencies in the time series data and the influence of external factors can be learned, and high-precision load forecasting can be provided.

[0028] S400: Invoking M groups of out-degree topological relationships of the M load distribution nodes in the power grid load topology, and constructing M load association prediction models of the M load distribution nodes according to the M groups of out-degree topological relationships.

[0029] Specifically, M groups of out-degree topological relationships are used for M load distribution nodes in the power grid load topology. In graph theory, out-degree represents the number of nodes a node points to. Out-degree topological relationships in the power grid represent the power flow relationship between each load distribution node and other nodes, particularly the load demand nodes affected by the load distribution node. Based on the out-degree topological relationships of the load distribution nodes, historical load data for the load demand nodes connected to the out-degree of the load distribution node are extracted from N historical load records. Based on the mapping relationship between the load distribution node and the load demand node, the historical load records of the load distribution node itself are extracted. A sliding time window is used to process the M historical load records and M groups of historical load records to obtain M instantaneous load distribution sets and M groups of instantaneous load demand sets. The data generated by the sliding window is used for multivariate regression analysis to establish a prediction model for the load distribution node. The load distribution node's distribution history is input and the load demand of the associated nodes is output. The load association prediction model describes the relationship between the output load of a load node and the demand of other load nodes, predicting how well the load distribution node's output meets the demand of its associated load nodes. A correlation prediction model is established for each load distribution node. By constructing a load correlation prediction model, the impact of the output load of the load distribution node on the demand of the associated load nodes can be accurately predicted.

[0030] S500: Map the K load single-point prediction models to the K load demand nodes of the power grid load topology, and map the M load association prediction models to the M load distribution nodes of the power grid load topology to complete the construction of the load multi-level graph calculation network.

[0031] Specifically, K trained single-point load prediction models are mapped to K load demand nodes in the power grid load topology. Single-point load prediction models are used to predict the load demand at a single load demand node. They are typically trained based on historical load data and external factors (such as weather and time), using time series prediction methods such as LSTM models (Long Short-Term Memory Networks). For each load demand node, the trained model is used to predict the load demand by inputting historical load data and external environmental factors (such as temperature, humidity, and time).

[0032] The M trained load association prediction models are mapped to the M load distribution nodes in the power grid's load topology. Predictions are made based on their out-degree topological relationships and associated load demands. The load association prediction model is used to predict the relationship between the output load of a load distribution node and other load nodes. Based on the power grid's topology and historical load data, it predicts how the output of a load distribution node affects other load demand nodes.

[0033] By mapping K single-point load prediction models and M associated load prediction models to corresponding nodes in the power grid load topology, a multi-level load graph computing network is ultimately formed. The power grid load topology is a directed graph that represents the various load nodes in the power grid and the power flow relationships between them. Nodes in the topology graph represent load points, and edges represent the direction of power flow. The multi-level load graph computing network is formed by mapping multiple load prediction models to the power grid load topology, forming a multi-level graph structure. In this structure, load nodes implement data flow and prediction calculations through graph computing models. Nodes at each level can perform load predictions independently or perform linked calculations based on the outputs of other nodes. By mapping load prediction models to different nodes in the power grid topology, the load demand and load distribution of each node can be accurately predicted, thereby improving the accuracy of power grid load predictions.

[0034] S600: acquiring the power supply load output by collecting K updated single-node time series data of the K load demand nodes in real time and synchronizing the K updated single-node time series data to the load multi-level graph calculation network for load forecasting cycle update.

[0035] Specifically, K updated single-node time series data from K load demand nodes are collected in real time, including weather data (temperature, humidity), time intervals, day of the week, and holiday status. The K updated single-node time series data are synchronized with K single-point load prediction models for prediction, resulting in K single-point load demands. Based on M groups of out-degree topological relationships and multi-layer node combinations, M load association prediction models are run to perform progressive load demand forecasts on the K single-point load demands, resulting in M ​​load distribution demands. The real-time data from the K load demand nodes is transmitted via a computing network to a multi-level load graph, updating the load forecast for each node. The multi-level load graph considers the spatiotemporal coupling between nodes and uses graph algorithms (such as graph convolutional networks and graph neural networks) to make load forecasts and scheduling decisions.

[0036] Progressive load demand forecasting is a prediction model based on known load demand. It takes the load demand of a specific node as input and progressively predicts the demand of other nodes (particularly load distribution nodes). The results of progressive load demand forecasting for K load demand nodes, i.e., M load distribution demands, are aggregated to produce the total load output of the power source. The load output of a power source refers to the amount of electricity output by a generation or power source node in the power grid, which must meet the load demand of the entire grid. By using K single-point load forecasting models and M load-related forecasting models, real-time load demand forecasting is achieved for each load node in the power grid, enabling optimization of grid dispatch strategies and ensuring a precise match between power supply and demand.

[0037] Furthermore, the present application S100 includes:

[0038] A grid node contribution analysis is performed on the target grid, and the N load nodes are obtained by screening based on the analysis results; N groups of node topology relationships are obtained by performing power flow path analysis on the N load nodes in the target grid; the N load nodes are connected according to the N groups of node topology relationships to complete the construction of the grid load topology.

[0039] Specifically, H grid nodes are extracted from the target power grid, which can be substations, power stations, regional distribution points, and other nodes with different load demands. The impact of each node on grid stability and load forecasting is assessed based on factors such as load fluctuation, fault propagation, and connectivity, identifying nodes with the greatest impact on grid operation. The contribution of each node is determined by analyzing its load fluctuation frequency, load fluctuation scale, fault radiation coverage, fault interval scale, node centrality, and node influence. From these H nodes, N load nodes with significant load fluctuations, high importance, or a significant impact on grid load scheduling are selected.

[0040] The power flow path analysis is performed on N load nodes in a target power grid. The power flow paths are simulated to determine how power flows from the power generation end to the load nodes and its transmission path within the power grid, focusing on the direction, magnitude, and potential bottlenecks of power flow. Node topology refers to the connectivity between nodes in a power grid. Node topology reflects how grid nodes are interconnected and how power flows between them. Based on the power flow paths and topological relationships of the N load nodes, these nodes are connected to complete the grid load topology. Taking into account the direction of power flow and the connectivity of each load node ensures that load forecasting and scheduling in the power grid more accurately reflect actual conditions. By screening load nodes with significant impact on grid operation and analyzing their power flow paths, critical nodes and potential bottlenecks in the power grid can be identified, reducing grid instability caused by load fluctuations or fault propagation and improving the grid's resilience to risks.

[0041] Furthermore, the present application further comprises the following steps:

[0042] Extract and obtain H grid nodes from the target grid; interactively obtain H historical load data and H historical operation and maintenance data of the H grid nodes; perform load fluctuation analysis based on the H historical load data to obtain H load fluctuation frequencies and H load fluctuation scales; perform fault impact analysis based on the H historical operation and maintenance data to obtain H fault radiation coverage rates and H fault interval scales; traverse the target grid using the H grid nodes to obtain H groups of hierarchical node connectivity, and obtain H node centralities by recursively calculating the H groups of hierarchical node connectivity; perform power flow calculation on the H grid nodes to obtain H node influences; and perform a comprehensive node contribution analysis on the H grid nodes based on the H load fluctuation frequencies, H load fluctuation scales, H fault radiation coverage rates, H fault interval scales, H node centralities, and H node influences to screen and obtain the N load nodes from the H grid nodes.

[0043] Specifically, H grid nodes are extracted from the target power grid, representing key facilities within the grid, such as substations and load points. Through collaboration with relevant departments or the grid data system, H historical load data and H historical operation and maintenance data are obtained for each of the H grid nodes. Historical load data refers to the power load data for each grid node over a certain period of time. This data typically shows the time-varying changes in parameters such as current, voltage, and power. This data is used to predict and analyze grid load fluctuations and demand trends. Historical operation and maintenance data refers to data related to the operation and maintenance of grid equipment, including equipment failure information, maintenance records, failure frequency, and repair times.

[0044] Based on H historical load data from H power grid nodes, load fluctuation analysis is performed to analyze the frequency (i.e., how often load changes occur) and the magnitude (i.e., the amplitude of load changes) of each node. For example, load data from a substation over a week shows the periodicity and amplitude of its load fluctuations, with larger fluctuations during peak hours on weekdays and more stable loads at night. Based on H historical operation and maintenance data from H power grid nodes, fault impact analysis is performed, including fault radiation coverage and fault interval scale. Fault radiation coverage describes the scope of the fault impact, i.e., the proportion of grid nodes affected when a fault occurs; the fault interval scale describes the interval between faults, i.e., the length of time between two faults.

[0045] The target power grid is traversed based on H power grid nodes to determine the hierarchical node connectivity of each node. The connectivity is recursively calculated based on the direct and indirect connections of the node. The centrality of each node evaluates the importance of the node in the power grid by calculating the degree of connection between the node and other nodes and its influence. Node connectivity represents the number of connections a node has with other nodes, reflecting the relative importance of the node in the power grid; centrality refers to the importance of a node in the network, which is usually quantified by calculating the direct and indirect connections of the node. For example, a node is directly connected to 3 power grid nodes, and then based on these 3 power grid nodes, it is indirectly connected to 7 power grid nodes to form the first set of hierarchical node connectivity. A set of recursive calculations of hierarchical node connectivity are performed based on the preset weight configuration of each level to obtain the centrality of the power grid node.

[0046] Perform power flow calculations on H power grid nodes to assess the influence of each node. Based on the power grid's topology and operating parameters, calculate parameters such as current and voltage for each node to determine the power flow path and load distribution. Obtain the H node influences, i.e., the range of influence of each node on power flow within the grid. A comprehensive analysis of the node contributions of the H power grid nodes is performed based on H load fluctuation frequencies, H load fluctuation scales, H fault radiation coverage rates, H fault interval scales, H node centralities, and H node influences. Determine the node contribution of each node's influence, i.e., the H node contributions.

[0047] H nodes are ranked by contribution from high to low, and the top N grid nodes are used as N load nodes. These nodes are crucial for load forecasting and grid scheduling. By comprehensively considering multiple factors such as load fluctuations, node connectivity, and the impact of faults, critical load nodes can be more accurately selected, thereby improving the overall accuracy of load forecasting.

[0048] Furthermore, the present application further comprises the following steps:

[0049] A node contribution evaluation function is pre-built, and the node contribution evaluation function is as follows: ;in, C D is the node contribution, P is the load fluctuation frequency, S is the load fluctuation scale, C is the fault radiation coverage, I is the fault interval scale, D is the node centrality, and F is the node influence; the H load fluctuation frequencies, H load fluctuation scales, H fault radiation coverage rates, H fault interval scales, H node centralities, and H node influences are synchronized to the node contribution evaluation function to obtain H node contributions; according to the serialized results of the H node contributions, the top N grid nodes are called as the N load nodes.

[0050] Specifically, a node contribution evaluation function is constructed to evaluate the contribution of grid nodes to grid load. The expression is: ;;in, C D Node contribution, which is used to quantify the contribution of each grid node; P is the load fluctuation frequency, which reflects the change of node load over time; S is the load fluctuation scale, which indicates the amplitude of load change; C Fault radiation coverage refers to the impact range of a fault on other nodes. I is the fault interval scale, the time interval between two faults; D is the node centrality, which is the importance of the grid node in the network, usually evaluated by calculating the connection relationship between the node and other nodes; F is the node influence, which is the influence range of the power flow of the node in the grid, usually obtained by power flow calculation, reflecting the control role of the node in the grid.

[0051] What is calculated is the impact of load fluctuation on the node contribution. 1+P is used to prevent division by zero errors when P is zero and also to smooth the nodes. What is calculated is the impact of fault radiation coverage on node contribution. An exponential decay factor is introduced to simulate the effect that the impact of failure intervals on node contribution gradually weakens over time. What is calculated is the impact of node influence on node contribution.

[0052] The H parameters of each grid node (load fluctuation frequency P, load fluctuation scale S, fault radiation coverage C, fault interval scale I, node centrality D, and node influence F) are input into the node contribution evaluation function to obtain the contribution of each grid node. After calculating the contribution of each node, the H node contributions are ranked from high to low. Based on the ranking results, the top N nodes are selected as key load nodes. These nodes are crucial for load forecasting and scheduling of the power grid. Through node contribution evaluation, the most critical load nodes for grid operation are identified, and load nodes with greater impact are screened for priority scheduling or real-time monitoring, thereby optimizing grid operation efficiency and resource allocation.

[0053] Furthermore, S300 of the present application includes: interactively obtaining a first historical load record and a first historical external data of a first load demand node; dividing the first historical external data into multiple sample single-node time series data based on a sliding time window, wherein the multiple sample single-node time series data have multiple sample last timestamp identifiers; obtaining multiple sample instantaneous load demands from the first historical load record according to the multiple sample last timestamps; using the multiple sample single-node time series data and the multiple sample instantaneous load demands as training data to train a first load single-point prediction model, wherein the first load single-point prediction model is an LSTM model; and so on, constructing K-1 load single-point prediction models for K-1 load demand nodes.

[0054] Specifically, a node is randomly selected from K load demand nodes as the first load demand node. A connection is established with the power grid's SCADA system, weather station, historical record database, and other systems to obtain the first historical load record and first historical external data. The first historical load record represents the historical load data of the target load demand node, recording the power demand of a node in the power grid over a specific time period, allowing analysis of the node's load trends and characteristics. Historical external data refers to external factors that affect load demand, such as weather data (temperature, humidity, etc.), time information (such as whether it is a weekday or a holiday), and other factors that may affect power demand. This data is external variables directly related to the load demand node and is used to predict future load demand.

[0055] Sliding time windows are a time series data processing method that segments time series data into multiple fixed-length samples. In this step, the sliding time window divides the first historical external data into multiple single-node time series data samples. Each single-node time series data sample contains continuous data within a specific time period, such as weather data or weekday information for a specific period, such as temperature, humidity, and holiday status for a specific time period (e.g., 9:00-11:00). The sample end timestamp is a time identifier for each sample data sample, indicating the end time of the data.

[0056] Based on the final timestamps of multiple samples, the corresponding instantaneous load demands for multiple samples are extracted from the first historical load record to ensure a one-to-one temporal correspondence between the time series data and the load demands. Instantaneous load demand refers to the power demand value of a grid load demand node at a specific point in time. Multiple samples of single-node time series data and the corresponding multiple samples of instantaneous load demands are used as training data and input into the LSTM model for training. Each set of data contains inputs (such as temperature, humidity, holidays, etc.) and outputs (such as the load demand at that moment). LSTM (Long Short-Term Memory) is a deep learning model particularly well-suited for processing time series data. It can capture long-term dependencies and process data with temporal order, especially data with temporal characteristics.

[0057] Preprocessing is performed on multiple samples of single-node time series data and multiple samples of instantaneous load demand, including normalization or standardization, to ensure that input features (such as temperature and humidity) are within the same scale range. This improves training efficiency and avoids numerical instability during training. An LSTM network is constructed and its structure determined. The core LSTM architecture consists of three gating mechanisms: a forget gate, an input gate, and an output gate. Each LSTM unit uses these gates to determine which information to retain and which to forget, and outputs a new state. The input layer inputs external environmental data (such as temperature and humidity), with each input sample associated with a timestamp. The LSTM layer learns the time series features of the input data through internal mechanisms. The model predicts the current load demand based on historical load changes and external factors. The output of the LSTM layer typically passes through one or more fully connected layers to convert the LSTM output into a target load forecast value. The final output layer provides the load forecast value for the grid node, typically a regression task, outputting the predicted load demand value.

[0058] In LSTM models, the training objective is to minimize the error (i.e., the loss function) between predicted and actual values. Mean squared error (MSE) is typically used as the loss function. Backpropagation is used to adjust the parameters (weights and biases) of the LSTM unit to reduce error. LSTM models typically require multiple iterations to achieve high accuracy. Each iteration is called a training epoch, and each training epoch involves performing both forward and backward propagation on all training examples. Multiple training epochs are typically set to ensure model convergence. During training, a portion of the dataset (the validation set) is often used to verify the model's generalization ability and avoid overfitting. The validation set data is not used in training and is used only to test the model's performance on unseen data. If the model's training error is small but the error on the validation set is large, this indicates overfitting and requires adjustment of model complexity or regularization.

[0059] Similarly, for the remaining K load demand nodes, the same method is used to construct single-point load forecasting models for each of the K-1 load demand nodes using historical and external data. Each prediction model for each load demand node is trained based on its historical load records and external data to accurately forecast load demand. Ultimately, K single-point load forecasting models corresponding to the K load demand nodes are obtained. By combining a sliding time window with an LSTM model, complex time series data can be processed to accurately predict future load demand for each load demand node. By accounting for the impact of external factors (such as weather and holidays), the forecast results are more consistent with actual conditions.

[0060] Furthermore, the present application S400 includes:

[0061] Interactively obtain N historical load records of the N load nodes; based on the M groups of out-degree topological relationships, call and obtain M groups of historical load records of M groups of associated load nodes connected to the out-degree of the M load distribution nodes from the N historical load records; based on the mapping relationship between the M load distribution nodes and the N load nodes, call the M historical load records of the M load distribution nodes from the N historical load records; use the sliding window length of the sliding time window to perform instantaneous record extraction on the M historical load records and the M groups of historical load records to obtain M instantaneous load distribution sets and M groups of instantaneous load demand sets; perform multivariate regression analysis on the M instantaneous load distribution sets and the M groups of instantaneous load demand sets to obtain M load association prediction functions; and construct the M load association prediction models based on the M load association prediction functions.

[0062] Specifically, N historical load records of N load nodes are interactively obtained, usually by connecting to the SCADA system, weather station, historical record database, etc. of the power grid. The historical load records represent the load demand data of the power grid nodes over a period of time. According to the M groups of out-degree topological relationships of the M load distribution nodes, the historical load records related to the M groups of associated load nodes connected to the out-degree of the M load distribution nodes are filtered out from the N historical load records. The out-degree indicates the number of other nodes that a node points to. The out-degree topological relationship of the load distribution node indicates how many other nodes the node can affect. The associated load node refers to the load demand node connected to the load distribution node through the out-degree topological relationship.

[0063] Based on the mapping relationship, the historical load records of the M load distribution nodes themselves are extracted from the N historical load records. The node mapping relationship represents the correspondence between load distribution nodes and load demand nodes. The historical load records of the M groups of associated load nodes are the historical load records of the M load demand nodes that are out-degree connected to the load distribution node. The historical load records of the M load distribution nodes are the load demand data of the load distribution nodes themselves.

[0064] Sliding time window is a time series analysis method that captures the changing trend of data by setting the window length and gradually moving the time window. In load forecasting, sliding time windows are usually used to process historical load records and divide them into multiple time periods for analysis. Using the sliding time window technology, instantaneous record extraction is performed on M historical load records and M groups of historical load records to obtain M instantaneous load distribution sets and M groups of instantaneous load demand sets. Instantaneous load refers to the load demand of the power grid at a specific time point. The load data extracted through the sliding time window obtains the load demand at the last moment, that is, the instantaneous load. Through the sliding window, the instantaneous load data of each load distribution node and load demand node are extracted separately. The instantaneous load data of the M load distribution nodes obtained constitute a set, and the instantaneous load data of the M load demand nodes also constitute another set.

[0065] Multiple regression analysis is a statistical method used to analyze the relationship between multiple independent variables and dependent variables. In power grid load forecasting, regression analysis can help model the mathematical relationship between load distribution nodes and their associated load nodes, and obtain M load association prediction functions. The extracted load data is used as input for multiple regression analysis to establish the relationship between load distribution nodes and associated load nodes. Through regression analysis, a set of prediction functions is obtained to predict the load demand of the target node. Based on the load association prediction function, the regression analysis results are integrated into the model. Each load distribution node has an independent load association prediction model, which can predict future load demand based on historical load data and associated load demand nodes. Through multiple regression analysis and historical load records extracted by sliding windows, the future load of the load distribution node is accurately predicted, while taking into account the influence of the load demand nodes associated with it.

[0066] Furthermore, S600 of the present application includes: synchronizing the K updated single-node time series data to the K load single-point prediction models of the load multi-level graph computing network to obtain K load single-point demands; hierarchically dividing the M load distribution nodes in the power grid load topology to obtain a multi-layer node combination; running the M load association prediction models to perform progressive load demand prediction on the K load single-point demands based on the M group out-degree topological relationship and the multi-layer node combination to obtain M load distribution demands; obtaining the power supply source load output of the source node from the M load distribution demands; when performing power supply scheduling of the source node with the power supply source load output as a constraint, starting the load scheduling of M-1 load distribution nodes in sequence based on the M-1 load distribution demands.

[0067] Specifically, K updated single-node time series data are synchronized to K load single-point prediction models for prediction, obtaining K load single-point demands. In the power grid load topology, the M load distribution nodes are divided into hierarchical levels to form a multi-layer node combination. Based on the topological structure of the power grid, the load distribution nodes are divided into different levels, and nodes at higher levels are responsible for a wider range of load scheduling. The load distribution nodes in the power grid are divided into levels based on the connection relationship between the load distribution nodes and other nodes. The role of each level node is different, and generally, the load distribution nodes at higher levels are responsible for more power distribution tasks.

[0068] Based on M groups of out-degree topological relationships and multi-layer node combinations, M load association prediction models are run to perform progressive predictions on the demands of K single load points, i.e., predictions are made from the load demand node layer by layer upward to the load distribution node, ultimately obtaining M load distribution demands. Load demand progressive prediction is a prediction model based on known load demand. It takes the load demand of a certain node as input and gradually predicts the demands of other nodes. By leveraging the out-degree topological relationships and hierarchical divisions in the power grid topology, the predicted values ​​of the bottom-level demand nodes are used as input data. Through a layer-by-layer transmission method, the predicted values ​​are input into the load association prediction model of the previous layer for load prediction. The prediction results obtained by the previous layer model serve as input data for the model of the next layer, and the prediction results are gradually transmitted until the power supply load output of the source node is obtained.

[0069] Based on the predicted load distribution requirements, the power output of the source node (e.g., a power station) is calculated and called upon. M load distribution requirements are aggregated to determine the power that the source node can or needs to output. The load output of the source node is crucial for power supply scheduling, ensuring that the power supply can meet the needs of each load distribution node. By calculating the load demand and the grid's carrying capacity, the power output that the source node should provide is determined.

[0070] The load output of the source node is a constraint on the power supply of the power grid. This constraint must not exceed the maximum load output capacity of the source node. In other words, the source node in the power grid can only provide a certain maximum power output, which must meet the total demand of all load distribution nodes and load demand nodes. Based on the load output of the source node, the power grid's load dispatcher begins adjusting the power demand of each load distribution node. Under the constraint of the source node's load output, load dispatching is initiated for each of the M-1 load distribution nodes. The source node first supplies power to nodes with higher priority or greater load demand, while subsequent nodes receive power in turn.

[0071] After power is exported from the source node, the load dispatch system sequentially dispatches M-1 load distribution nodes according to the load distribution requirements. The power demand of each node is gradually met until the demand of the last node is met. The order in which dispatch is initiated is typically based on the importance of the load distribution nodes, the amount of demand, or other factors such as priority and regional urgency. The power grid adjusts the flow of power based on the load distribution requirements to ensure a balance between power supply and demand. Power is distributed to each load distribution node in a sequential and gradual manner, taking into account factors such as power transmission paths, node priority, and load demand. If power supply capacity allows, load dispatch is initiated sequentially according to node priority.

[0072] If power shortages occur during load distribution (for example, if demand at certain nodes exceeds the maximum output of the source node), the grid dispatcher will readjust the dispatch of each load distribution node based on power supply constraints. Using K single-point load prediction models and M load-related prediction models, real-time load demand forecasts are achieved for each load node in the grid, optimizing grid dispatch strategies and ensuring a precise match between power supply and demand.

[0073] In summary, the graph computing-based smart grid load forecasting method provided by this application has the following technical effects:

[0074] By performing load analysis on the target power grid, a power grid load topology is constructed; an in-degree connection relationship analysis is performed on N load nodes in the power grid load topology, and the N load nodes are divided into K load demand nodes and M load distribution nodes, where K+M=N, K>M; K load single-point prediction models of the K load demand nodes are pre-constructed; M groups of out-degree topological relationships of the M load distribution nodes are called in the power grid load topology, and M load association prediction models of the M load distribution nodes are constructed according to the M groups of out-degree topological relationships; the K load single-point prediction models are mapped and loaded to the K load demand nodes of the power grid load topology, and the M load association prediction models are mapped and loaded to the M load distribution nodes of the power grid load topology, thereby completing the construction of a load multi-level graph computing network; K updated single-node time series data of the K load demand nodes are collected in real time, and the K updated single-node time series data are synchronized to the load multi-level graph computing network for load prediction cycle update, thereby obtaining the power supply load output. That is to say, by constructing the power grid load topology, the load transfer relationship between nodes is converted into the in-degree and out-degree relationship in the graph structure. For the demand node, a single-point load prediction model is constructed to capture local load changes; for the distribution node, its out-degree relationship is used to construct a load association prediction model to accurately characterize the load transfer law between nodes, avoid the deviation caused by the independence assumption of the single-point prediction model, and improve the accuracy of load prediction.

[0075] Example 2: Based on the same inventive concept as the graph computing smart grid load forecasting method in the above-mentioned Example 1, this application also provides a graph computing smart grid load forecasting system, please refer to the attached Figure 2 , the graph computing smart grid load forecasting system includes:

[0076] A load analysis module 11 is used to construct a power grid load topology by performing load analysis on the target power grid; a connection relationship analysis module 12 is used to perform in-degree connection relationship analysis on N load nodes in the power grid load topology, and divide the N load nodes into K load demand nodes and M load distribution nodes, wherein K+M=N, K>M; a first model construction module 13 is used to pre-construct K load single-point prediction models of the K load demand nodes; a second model construction module 14 is used to call the M groups of out-degree topological relationships of the M load distribution nodes in the power grid load topology, and according to the The M groups of out-degree topological relationships construct M load association prediction models of the M load distribution nodes; a graph computing network construction module 15, the graph computing network construction module 15 is used to map and load the K load single-point prediction models to the K load demand nodes of the power grid load topology, and map and load the M load association prediction models to the M load distribution nodes of the power grid load topology, to complete the construction of the load multi-level graph computing network; a load prediction module 16, the load prediction module 16 is used to collect K updated single-node time series data of the K load demand nodes in real time, and synchronize the K updated single-node time series data to the load multi-level graph computing network for load prediction cycle update to obtain the power supply load output.

[0077] Furthermore, the load analysis module 11 in the graph computing smart grid load forecasting system is further configured to:

[0078] A grid node contribution analysis is performed on the target grid, and the N load nodes are obtained by screening based on the analysis results; N groups of node topology relationships are obtained by performing power flow path analysis on the N load nodes in the target grid; the N load nodes are connected according to the N groups of node topology relationships to complete the construction of the grid load topology.

[0079] Furthermore, the load analysis module 11 in the graph computing smart grid load forecasting system is further configured to:

[0080] Extract and obtain H grid nodes from the target grid; interactively obtain H historical load data and H historical operation and maintenance data of the H grid nodes; perform load fluctuation analysis based on the H historical load data to obtain H load fluctuation frequencies and H load fluctuation scales; perform fault impact analysis based on the H historical operation and maintenance data to obtain H fault radiation coverage rates and H fault interval scales; traverse the target grid using the H grid nodes to obtain H groups of hierarchical node connectivity, and obtain H node centralities by recursively calculating the H groups of hierarchical node connectivity; perform power flow calculation on the H grid nodes to obtain H node influences; and perform a comprehensive node contribution analysis on the H grid nodes based on the H load fluctuation frequencies, H load fluctuation scales, H fault radiation coverage rates, H fault interval scales, H node centralities, and H node influences to screen and obtain the N load nodes from the H grid nodes.

[0081] Furthermore, the load analysis module 11 in the graph computing smart grid load forecasting system is further configured to:

[0082] A node contribution evaluation function is pre-built, and the node contribution evaluation function is as follows: ; Wherein, CD is the node contribution, P is the load fluctuation frequency, S is the load fluctuation scale, C is the fault radiation coverage, I is the fault interval scale, D is the node centrality, and F is the node influence; the H load fluctuation frequencies, H load fluctuation scales, H fault radiation coverage, H fault interval scales, H node centralities and H node influences are synchronized to the node contribution evaluation function to obtain H node contributions; according to the serialization results of the H node contributions, the top N grid nodes are called as the N load nodes.

[0083] Furthermore, the first model building module 13 in the graph computing smart grid load forecasting system is further configured to:

[0084] Interactively obtain a first historical load record and first historical external data of a first load demand node; divide the first historical external data into multiple sample single-node time series data based on a sliding time window, wherein the multiple sample single-node time series data have multiple sample last timestamp identifiers; obtain multiple sample instantaneous load demands from the first historical load record according to the multiple sample last timestamps; use the multiple sample single-node time series data and the multiple sample instantaneous load demands as training data to train a first load single-point prediction model, wherein the first load single-point prediction model is an LSTM model; and so on, construct K-1 load single-point prediction models for K-1 load demand nodes.

[0085] Furthermore, the second model building module 14 in the graph computing smart grid load forecasting system is further configured to:

[0086] Interactively obtain N historical load records of the N load nodes; based on the M groups of out-degree topological relationships, call and obtain M groups of historical load records of M groups of associated load nodes connected to the out-degree of the M load distribution nodes from the N historical load records; based on the mapping relationship between the M load distribution nodes and the N load nodes, call the M historical load records of the M load distribution nodes from the N historical load records; use the sliding window length of the sliding time window to perform instantaneous record extraction on the M historical load records and the M groups of historical load records to obtain M instantaneous load distribution sets and M groups of instantaneous load demand sets; perform multivariate regression analysis on the M instantaneous load distribution sets and the M groups of instantaneous load demand sets to obtain M load association prediction functions; and construct the M load association prediction models based on the M load association prediction functions.

[0087] Furthermore, the load forecasting module 16 in the graph computing smart grid load forecasting system is further configured to:

[0088] The K updated single-node time series data are synchronized to the K load single-point prediction models of the load multi-level graph computing network to obtain K load single-point demands; the M load distribution nodes are hierarchically divided in the power grid load topology to obtain a multi-layer node combination; based on the M group out-degree topological relationship and the multi-layer node combination, the M load association prediction models are run to perform progressive load demand prediction on the K load single-point demands to obtain M load distribution demands; the power supply source load output of the source node is obtained from the M load distribution demands; when the power supply scheduling of the source node is performed with the power supply source load output as a constraint, the load scheduling of the M-1 load distribution nodes is started in sequence based on the M-1 load distribution demands.

[0089] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. Figure 1 The graph-based smart grid load forecasting method and specific examples in Example 1 are also applicable to the graph-based smart grid load forecasting system in this embodiment. The detailed description of the graph-based smart grid load forecasting method above will clearly explain the graph-based smart grid load forecasting system in this embodiment. For the sake of brevity, a detailed description will not be given here. The system disclosed in this embodiment corresponds to the method disclosed in this embodiment, so the description is relatively brief. For relevant details, refer to the method description.

[0090] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present application. Various modifications to these embodiments will be readily apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application is not limited to the embodiments shown herein, but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

[0091] Obviously, those skilled in the art may make various changes and modifications to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the present application and its equivalents, the present application is intended to include these modifications and variations.

Claims

1. A graph computing-based smart grid load forecasting method, characterized in that: include: By analyzing the load of the target power grid, the power grid load topology is constructed; Performing an in-degree connection relationship analysis on N load nodes in the power grid load topology, dividing the N load nodes into K load demand nodes and M load distribution nodes, where K+M=N, K>M; Pre-building K load single-point prediction models for the K load demand nodes; Invoking M groups of out-degree topological relationships of the M load distribution nodes in the power grid load topology, and constructing M load association prediction models of the M load distribution nodes according to the M groups of out-degree topological relationships; Mapping and loading the K load single-point prediction models to the K load demand nodes of the power grid load topology, and mapping and loading the M load association prediction models to the M load distribution nodes of the power grid load topology, thereby completing the construction of the load multi-level graph calculation network; The power supply load output is obtained by collecting K updated single-node time series data of the K load demand nodes in real time and synchronizing the K updated single-node time series data to the load multi-level graph calculation network for load forecasting cycle update.

2. The method for smart grid load forecasting based on graph computing according to claim 1, wherein: Pre-building K load single-point prediction models for the K load demand nodes, the method comprising: interactively obtaining a first historical load record and first historical external data of a first load demand node; Dividing the first historical external data into a plurality of sample single-node time series data based on a sliding time window, wherein the plurality of sample single-node time series data have a plurality of sample last timestamp identifiers; Obtaining a plurality of sample instantaneous load demands from the first historical load record according to the last timestamps of the plurality of samples; Using the plurality of sample single-node time series data and the plurality of sample instantaneous load demands as training data, training a first load single-point prediction model, wherein the first load single-point prediction model is an LSTM model; Similarly, K-1 load single-point prediction models for K-1 load demand nodes are constructed.

3. The graph computing-based smart grid load forecasting method according to claim 2, wherein: Constructing M load association prediction models for the M load distribution nodes according to the M groups of out-degree topological relationships, the method comprising: Interactively obtaining N historical load records of the N load nodes; According to the M groups of out-degree topological relationships, from the N historical load records, M groups of historical load records of M groups of associated load nodes connected with the out-degree of the M load distribution nodes are retrieved; According to the mapping relationship between the M load distribution nodes and the N load nodes, calling the M historical load records of the M load distribution nodes from the N historical load records; Performing instantaneous record extraction on the M historical load records and the M groups of historical load records using the sliding window length of the sliding time window to obtain M instantaneous load distribution sets and M groups of instantaneous load demand sets; Performing a multiple regression analysis on the M instantaneous load distribution sets and the M groups of instantaneous load demand sets to obtain M load correlation prediction functions; The M load correlation prediction models are constructed based on the M load correlation prediction functions.

4. The graph computing-based smart grid load forecasting method according to claim 3, wherein: The method includes: acquiring K updated single-node time series data of the K load demand nodes in real time, synchronizing the K updated single-node time series data to the load multi-level graph computing network for load forecasting cyclic update, and obtaining a power supply load output. Synchronizing the K updated single-node time series data to the K load single-point prediction models of the load multi-level graph calculation network to obtain K load single-point demands; Dividing the M load distribution nodes into different levels in the power grid load topology to obtain a multi-layer node combination; According to the M groups of out-degree topological relationships and multi-layer node combinations, the M load association prediction models are run to perform progressive load demand prediction on the K single-point load demands to obtain M load distribution demands; Obtain the power supply load output of the source node from the M load distribution demand calls; When power supply scheduling of the source node is performed with the power supply source load output as a constraint, load scheduling of M-1 load distribution nodes is started in sequence based on M-1 load distribution requirements.

5. The method for smart grid load forecasting based on graph computing according to claim 1, wherein: By performing load analysis on a target power grid and constructing a power grid load topology, the method includes: Performing grid node contribution analysis on the target grid, and screening and obtaining the N load nodes according to the analysis results; Obtaining N groups of node topology relationships by performing power flow path analysis on the N load nodes in the target power grid; The N load nodes are connected according to the N groups of node topological relationships to complete the construction of the power grid load topology.

6. The graph computing-based smart grid load forecasting method according to claim 5, wherein: Performing grid node contribution analysis on the target grid and screening the N load nodes based on the analysis results, the method comprising: Extracting and obtaining H power grid nodes from the target power grid; Interactively obtaining H historical load data and H historical operation and maintenance data of the H power grid nodes; Performing load fluctuation analysis based on the H historical load data to obtain H load fluctuation frequencies and H load fluctuation scales; Perform fault impact analysis based on the H historical operation and maintenance data to obtain H fault radiation coverage rates and H fault interval scales; Using the H power grid nodes to traverse the target power grid, obtain H groups of hierarchical node connectivity, and obtain H node centralities by recursively calculating the H groups of hierarchical node connectivity; Performing power flow calculation on the H power grid nodes to obtain the influence of the H nodes; Based on the H load fluctuation frequencies, H load fluctuation scales, H fault radiation coverage rates, H fault interval scales, H node centralities, and H node influences, a comprehensive node contribution analysis is performed on the H power grid nodes to screen and obtain the N load nodes from the H power grid nodes.

7. The method for smart grid load forecasting based on graph computing according to claim 6, characterized in that: Based on the H load fluctuation frequencies, H load fluctuation scales, H fault radiation coverage rates, H fault interval scales, H node centralities, and H node influences, a comprehensive node contribution analysis is performed on the H power grid nodes to screen and obtain the N load nodes from the H power grid nodes. The method includes: A node contribution evaluation function is pre-built, and the node contribution evaluation function is as follows: ;in, C D is the node contribution, P is the load fluctuation frequency, S is the load fluctuation scale, C is the fault radiation coverage, I is the fault interval scale, D is the node centrality, F is the node influence; The H load fluctuation frequencies, H load fluctuation scales, H fault radiation coverage rates, H fault interval scales, H node centralities, and H node influences are synchronized with the node contribution evaluation function to obtain H node contributions; and based on the serialized results of the H node contributions, the top N grid nodes are called as the N load nodes.

8. A graph computing-based smart grid load forecasting system, characterized in that: The steps for implementing the graph computing smart grid load forecasting method according to any one of claims 1 to 7, wherein the graph computing smart grid load forecasting system comprises: A load analysis module, configured to perform load analysis on a target power grid and construct a power grid load topology; A connection relationship analysis module, wherein the connection relationship analysis module is used to perform in-degree connection relationship analysis on N load nodes in the power grid load topology, and divide the N load nodes into K load demand nodes and M load distribution nodes, wherein K+M=N, K>M; A first model building module, wherein the first model building module is used to pre-build K load single-point prediction models of the K load demand nodes; A second model building module, the second model building module is used to call M groups of out-degree topological relationships of the M load distribution nodes in the power grid load topology, and build M load association prediction models of the M load distribution nodes according to the M groups of out-degree topological relationships; A graph computing network construction module, which is used to map and load the K load single-point prediction models to the K load demand nodes of the power grid load topology, and map and load the M load association prediction models to the M load distribution nodes of the power grid load topology, thereby completing the construction of a multi-level load graph computing network; The load forecasting module is used to collect K updated single-node time series data of the K load demand nodes in real time, and synchronize the K updated single-node time series data to the load multi-level graph calculation network for load forecasting cycle update to obtain the power supply load output.

Images