An urban power load prediction method and device, electronic equipment and storage medium

Abstract

This application belongs to the field of power load forecasting technology, specifically disclosing a method, device, electronic device, and storage medium for urban power load forecasting. The method includes: dividing a regional map of the city to be forecasted into a regional map structure centered on multiple regional nodes according to a preset regional range; determining a target map propagation matrix based on the Euclidean distance, electrical topology information, and other multi-source correlation information between each regional node; determining target time-series characteristic information based on the power load data, meteorological and radiation data, comprehensive electricity consumption information of personnel, and photovoltaic power generation installed capacity density of each regional node in the most recent historical period; and inputting the target map propagation matrix and target time-series characteristic information into a power load forecasting model to obtain the power load forecasting result of the city to be forecasted, output by the power load forecasting model. This application significantly improves the accuracy and precision of urban power load forecasting.

Description

Technical Field

[0001] This application belongs to the field of power load forecasting technology, specifically to the field of machine learning technology, and more specifically, to a method, device, electronic device and storage medium for urban power load forecasting. Background Technology

[0002] With the rapid development of urbanization and new energy sources, the installed capacity of distributed photovoltaic (PV) systems on residential and commercial rooftops continues to grow, and urban electricity consumption is shifting from being dominated by "total load" to "net load." Under clear conditions, PV output experiences a significant "dip" at midday, while net load rebounds rapidly at sunset, exhibiting typical phenomena such as the "duck curve." Cloud movement can also cause synchronous fluctuations and abrupt changes in net load in adjacent areas. This places higher demands on intraday dispatching, energy storage coordination, and the safe operation of the distribution network, making urban power load forecasting crucial.

[0003] In existing technologies, urban power load forecasting often uses single-point time series methods to independently model the load of each region, which makes it difficult to characterize cross-regional propagation and synchronous fluctuations. There are also graph learning-based schemes, but they often only use a single adjacency relationship for propagation, ignoring the impact of various urban operating mechanisms on power load, resulting in low accuracy of the final urban power load forecast.

[0004] Therefore, how to better predict urban power load has become a technical problem that the industry urgently needs to solve. Summary of the Invention

[0005] In view of the shortcomings of the existing technology, the purpose of this application is to better realize the prediction of urban power load, and to solve the problem of low prediction accuracy of existing urban power load prediction schemes.

[0006] To achieve the above objectives, in a first aspect, this application provides a method for predicting urban power load, comprising: dividing the regional map of the city to be predicted into a regional map structure centered on multiple regional nodes according to a preset regional range;

[0007] The target graph propagation matrix is ​​determined based on the Euclidean distance between each of the regional nodes, electrical topology information, and other multi-source association information; the other multi-source association information includes at least one of the following: shortest path length, functional similarity, traffic flow, and irradiance information.

[0008] Based on the power load data, meteorological and radiation data, comprehensive electricity consumption information of personnel, and photovoltaic power generation installed capacity density of each regional node in the most recent historical period, target time-series characteristic information is determined.

[0009] The target graph propagation matrix and the target time series feature information are input into the power load prediction model to obtain the power load prediction result of the city to be tested output by the power load prediction model.

[0010] The power load prediction model is trained based on target training data samples and corresponding power load data labels; the target training data samples include target graph propagation matrix samples and target time series feature information samples.

[0011] Optionally, the other multi-source association information includes shortest path length, functional similarity, traffic flow, and irradiance information; determining the target graph propagation matrix based on the Euclidean distance between each of the regional nodes, electrical topology information, and other multi-source association information includes:

[0012] Based on the Euclidean distance between each of the said region nodes, determine the spatial proximity matrix that associates all said region nodes;

[0013] Based on the electrical topology information between each of the said regional nodes, an electrical topology matrix relating all said regional nodes is determined;

[0014] Based on the shortest path length between each of the said regional nodes, determine the road network connectivity matrix that associates all said regional nodes;

[0015] Based on the functional similarity between each of the said regional nodes, a functional similarity matrix is ​​determined that associates all said regional nodes.

[0016] Based on the traffic flow between each of the said regional nodes, a passenger flow association matrix is ​​determined that relates all said regional nodes;

[0017] Based on the irradiation information of each of the said regional nodes, an irradiation correlation matrix relating all said regional nodes is determined;

[0018] The target map propagation matrix is ​​determined based on the spatial proximity matrix, the road network connectivity matrix, the functional similarity matrix, the pedestrian flow correlation matrix, the electrical topology matrix, and the irradiance correlation matrix.

[0019] Optionally, determining the target map propagation matrix based on the spatial proximity matrix, the road network connectivity matrix, the functional similarity matrix, the pedestrian flow correlation matrix, the electrical topology matrix, and the irradiance correlation matrix includes:

[0020] The fusion adjacency matrix is ​​determined by weighting the spatial proximity matrix, the road network connectivity matrix, the functional similarity matrix, the pedestrian flow correlation matrix, the electrical topology matrix, and the irradiance correlation matrix, as well as the corresponding preset learnable weight parameter group.

[0021] Based on two pre-set sets of learnable node embedding parameters, an adaptive adjacency matrix is ​​determined.

[0022] The target graph propagation matrix is ​​determined based on the fused adjacency matrix and the adaptive adjacency matrix.

[0023] Optionally, determining the spatial proximity matrix of all associated regional nodes based on the Euclidean distance between each of the regional nodes includes:

[0024] The Gaussian kernel weight matrix is ​​obtained by calculating the Euclidean distance between each node in the region using a Gaussian kernel weight function.

[0025] The spatial proximity matrix is ​​obtained by sparsifying the elements of each row in the Gaussian kernel weight matrix.

[0026] Optionally, determining the functional similarity matrix associated with all the regional nodes based on the functional similarity between each of the regional nodes includes:

[0027] Based on the POI data of each of the said regional nodes, a multidimensional POI ratio vector for each of the said regional nodes is determined; the POI data is determined based on the number of points of interest in the social function categories possessed by each of the said regional nodes;

[0028] The functional similarity matrix is ​​determined by calculating the similarity between pairs of nodes using the multidimensional POI ratio vector of each region node.

[0029] Optionally, determining the electrical topology matrix associated with all the regional nodes based on the electrical topology information between each of the regional nodes includes:

[0030] Based on the electrical topology information between each of the regional nodes, determine the shortest electrical path and indication information of the same feeder between each of the regional nodes;

[0031] The electrical topology matrix is ​​obtained by mapping the shortest electrical path and the indication information of the same feeder between each of the regional nodes into a weight matrix between the nodes.

[0032] Optionally, the loss function used in the power load forecasting model during model training is:

[0033] ;

[0034] In the formula, This represents the power load forecast information for all regional nodes; This represents the actual power load information for all regional nodes; This indicates the spatial proximity information between regional nodes; This indicates the differences in net power load between regional nodes; These represent preset coefficients; This represents the mean absolute error function; This represents the root mean square error function.

[0035] Secondly, this application provides an urban power load forecasting device, comprising:

[0036] The partitioning module is used to divide the regional map of the city to be tested into a regional map structure centered on multiple regional nodes according to a preset regional range.

[0037] The first processing module is used to determine the target graph propagation matrix based on the Euclidean distance between each of the regional nodes, electrical topology information, and other multi-source association information; the other multi-source association information includes at least one of the following: shortest path length, functional similarity, traffic flow, and irradiance information.

[0038] The second processing module is used to determine target time-series characteristic information based on the power load data, meteorological and radiation data, comprehensive electricity consumption information of personnel and photovoltaic power generation installed capacity of each regional node in the most recent historical period.

[0039] The prediction module is used to input the target graph propagation matrix and the target time series feature information into the power load prediction model to obtain the power load prediction result of the city to be tested output by the power load prediction model;

[0040] The power load prediction model is trained based on target training data samples and corresponding power load data labels; the target training data samples include target graph propagation matrix samples and target time series feature information samples.

[0041] Thirdly, this application provides an electronic device, comprising: at least one memory for storing a program; and at least one processor for executing the program stored in the memory, wherein when the program stored in the memory is executed, the processor is configured to execute the method described in the first aspect or any possible implementation thereof.

[0042] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when run on a processor, causes the processor to perform the method described in the first aspect or any possible implementation thereof.

[0043] Fifthly, this application provides a computer program product that, when run on a processor, causes the processor to perform the method described in the first aspect or any possible implementation thereof.

[0044] It is understood that the beneficial effects of the second to fifth aspects mentioned above can be found in the relevant descriptions in the first aspect mentioned above, and will not be repeated here.

[0045] Overall, the technical solutions conceived in this application have the following beneficial effects compared with the prior art:

[0046] This application provides a method, device, electronic device, and storage medium for predicting urban power load. By dividing the city area map into nodes, and utilizing Euclidean distance, electrical topology information, shortest path length, functional similarity, traffic flow, and solar irradiance information between the regional nodes, it simultaneously constructs multiple types of multi-source adjacency relationships on the same set of regional nodes. These relationships can include spatial proximity, electrical topology, road network connectivity, functional similarity, pedestrian flow correlation, and irradiance correlation. This data is then combined with time-series feature data constructed from multi-source observations that change over time to form multi-source spatiotemporal map data. This data is then used by a power load prediction model trained by a spatiotemporal map neural network for identification and prediction. This significantly enhances the model's generalization ability and robustness in predicting power load under complex urban scenarios and abnormal days (such as holidays, event days, and extreme weather days), greatly improving the accuracy and precision of urban power load prediction. Attached Figure Description

[0047] Figure 1 This is a flowchart illustrating the urban power load forecasting method provided in an embodiment of this application;

[0048] Figure 2 This is one of the schematic diagrams of the multi-relationship city map in the urban power load forecasting method provided in the embodiments of this application;

[0049] Figure 3 This is the second schematic diagram of the multi-relationship city map form in the urban power load forecasting method provided in the embodiments of this application;

[0050] Figure 4 This is the third schematic diagram of the multi-relationship city map form in the urban power load forecasting method provided in the embodiments of this application;

[0051] Figure 5 This is one of the schematic diagrams illustrating the prediction results of the urban power load prediction method provided in the embodiments of this application;

[0052] Figure 6 This is the second schematic diagram of the prediction results of the urban power load prediction method provided in the embodiments of this application;

[0053] Figure 7 This is a schematic diagram of the urban power load forecasting device provided in the embodiments of this application;

[0054] Figure 8 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0055] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0056] 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 the terms "exemplary" or "for example" is intended to present the relevant concepts in a specific manner.

[0057] The embodiments of this application are described below with reference to the accompanying drawings.

[0058] The inventors' research revealed that net electricity load at the urban scale is coupled with multiple factors: first, temporal factors, such as temperature, humidity, wind speed, cloud cover, holidays, and time-of-use pricing; second, spatial factors, such as spatial proximity of adjacent areas, road accessibility, similarity of functional layout, and population travel intensity; third, electrical factors, such as feeder-substation topology and local permeability differences; and fourth, photoelectric factors, such as spatial correlation of irradiance, roof suitability, component orientation / tilt angle, and shading conditions. These factors act simultaneously in both time and space, resulting in strong spatiotemporal correlation and cross-regional propagation characteristics in the net load. Therefore, this application proposes the following method for predicting urban electricity load.

[0059] Figure 1 This is a flowchart illustrating the urban power load forecasting method provided in this application embodiment, as shown below. Figure 1 As shown, it includes:

[0060] Step S1: Divide the regional map of the city to be tested into a regional map structure centered on multiple regional nodes according to the preset regional range.

[0061] Step S2: Determine the target graph propagation matrix based on the Euclidean distance between each regional node, electrical topology information, and other multi-source association information; other multi-source association information includes at least one of the following: shortest path length, functional similarity, traffic flow, and irradiance information.

[0062] Step S3: Based on the power load data, meteorological and radiation data, comprehensive electricity consumption information of personnel, and photovoltaic power generation installed capacity density of each regional node in the most recent historical period, determine the target time series characteristic information;

[0063] Step S4: Input the target graph propagation matrix and target time series feature information into the power load prediction model to obtain the power load prediction results of the city to be tested output by the power load prediction model.

[0064] The power load forecasting model is trained based on the target training data samples and the corresponding power load data labels; the target training data samples include target graph propagation matrix samples and target time series feature information samples.

[0065] Specifically, the electrical topology information described in the embodiments of this application refers to the topological distance between each regional node on the electrical topology diagram from the distribution feeder to the substation, as well as feeder indication information, etc.

[0066] Other multi-source association information described in the embodiments of this application includes at least one of the following: shortest path length, functional similarity, traffic flow, and irradiance information.

[0067] Here, the shortest path length refers to the shortest path distance between each regional node on the road map; functional similarity refers to the similarity of the proportion of each regional node in the urban area corresponding to multiple points of interest (POI) functional categories (such as residential, commercial, industrial, educational, medical, transportation hub, public administration, etc.); traffic flow refers to the OD flow data between regional nodes; and irradiance information refers to the spatial correlation of solar irradiance or cloud movement between regional nodes at the same time.

[0068] The comprehensive electricity consumption information for personnel described in this application embodiment may specifically include information such as holiday indications, time-of-use electricity prices, and population density indices.

[0069] The target graph propagation matrix described in the embodiments of this application represents the graph structure data used to extract the spatial relationships between nodes when inputting into the power load prediction model.

[0070] The target time-series feature information described in the embodiments of this application represents the time dynamic data used to input the power load prediction model for node time-series feature extraction.

[0071] The power load prediction model described in this application refers to a model built based on a spatiotemporal graph neural network, which can be trained based on target training data samples and corresponding power load data labels. Here, the target training data samples specifically include target graph propagation matrix samples and target time-series feature information samples.

[0072] Understandably, the target map propagation matrix sample is determined based on data samples composed of Euclidean distances between nodes in each region, electrical topology information, and other multi-source correlation information over historical periods; the target time-series feature information sample is determined based on power load data, meteorological and radiation data, comprehensive electricity consumption information of personnel, and photovoltaic power generation installed capacity density of each node in each region over historical periods.

[0073] The power load forecasting results described in this application embodiment can specifically be net power load information, which is equal to the difference between the total power load and the photovoltaic grid-connected power load.

[0074] In the embodiments of this application, in step S1, the regional map of the city to be tested is first divided into several non-overlapping partitions according to a preset regional range. Each partition corresponds to a node in the graph, thereby discretizing the city region into fixed-type statistical units and mapping them as graph nodes, resulting in a regional graph structure centered on multiple regional nodes.

[0075] The preset area can be determined during implementation by fixing it as a hexagonal grid of equal area, a street / community, a functional area, or a power distribution area aggregation, and a unique node ID can be generated for each unit. Preferably, a hexagonal grid of equal area is used (e.g., the side length can range from 250m to 1000m), with the grid endpoints serving as area nodes to ensure full coverage and scale consistency. Simultaneously, static attributes can be recorded for each node, including building and functional attributes, road connectivity indicators, population and employment intensity, as well as photovoltaic (PV) installation density, roof suitability rate, and typical component orientation / tilt distribution, etc.

[0076] It should be noted that data collection and standardization can be performed before step S1. Specifically, the data listed in Table 1 should be collected within the administrative boundaries of the target city; all time series should be standardized to a time granularity of 15 minutes, 30 minutes, or 60 minutes, and spatial data should be standardized to the same plane coordinate system; missing or abnormal time data should be repaired by combining analysis and interpolation algorithms with similar daily data; non-numerical items should be standardized and aggregated into a two-dimensional table of "node × time" according to the node index.

[0077] Furthermore, a static attribute vector can be constructed for each region node, with fields and definitions shown in Table 2; static attributes serve as the foundation for cold start and interpretable analysis during the modeling phase and do not change over time.

[0078] Table 1

[0079]

[0080] Table 2

[0081]

[0082] In the embodiments of this application, in step S2, a multi-relationship adjacency matrix is ​​constructed and weighted fusion is performed based on the Euclidean distance between each region node, electrical topology information, and other multi-source association information to determine the target graph propagation matrix. Among them, other multi-source association information includes at least one of the following: shortest path length, functional similarity, traffic flow, and irradiance information. That is to say, other multi-source association information can be a single parameter of shortest path length, functional similarity, traffic flow, and irradiance information, or it can be a combination of multiple parameters of the four parameters.

[0083] Based on the above embodiments, as an optional embodiment, other multi-source association information includes shortest path length, functional similarity, traffic flow, and irradiance information; based on the Euclidean distance between each regional node, electrical topology information, and other multi-source association information, the target graph propagation matrix is ​​determined, including:

[0084] Based on the Euclidean distance between each region node, determine the spatial proximity matrix that associates all region nodes;

[0085] Based on the electrical topology information between each region node, determine the electrical topology matrix that associates all region nodes;

[0086] Based on the shortest path length between each regional node, determine the road network connectivity matrix that connects all regional nodes;

[0087] Based on the functional similarity between each region node, determine the functional similarity matrix of all associated region nodes;

[0088] Based on the passenger flow between each regional node, determine the passenger flow correlation matrix that connects all regional nodes;

[0089] Based on the irradiance information of each regional node, determine the irradiance correlation matrix that associates all regional nodes;

[0090] The target map propagation matrix is ​​determined based on the spatial proximity matrix, road network connectivity matrix, functional similarity matrix, pedestrian flow correlation matrix, electrical topology matrix, and irradiance correlation matrix.

[0091] Specifically, in the embodiments of this application, other multi-source association information may simultaneously include four types of parameters: shortest path length, functional similarity, traffic flow, and irradiance information, thereby constructing an adjacency set of six types of relationships on the same regional node. .

[0092] Among them, six types of adjacency matrices can be constructed based on the Euclidean distance, electrical topology information, shortest path length, functional similarity, traffic flow and irradiance information between each regional node. The list of urban graph relationship matrices that can be represented by these six types of adjacency matrices can be seen in Table 3.

[0093] Table 3

[0094]

[0095] Here, gate strength can be understood as the passenger flow count recorded by public transportation facilities within a given time step, such as the number of times a subway station, light rail station, or bus station swipes its card to enter or exit within that time step.

[0096] In the embodiments of this application, these passenger flow counts can be mapped to their respective regional nodes or regional node pairs according to routes and stations, in order to characterize the strength of passenger flow between different regions for public transportation, thereby constructing a passenger flow correlation matrix.

[0097] The method in this application embodiment constructs six types of relationships on the same node set simultaneously: spatial proximity, road network connectivity, functional similarity, pedestrian flow association, electrical topology, and radiation correlation. Based on these six types of relationships, a graph structure data is constructed. By integrating multi-source data and multi-dimensional information, the key factors affecting changes in urban power load can be captured more comprehensively, thereby improving the accuracy and robustness of urban power load forecasting.

[0098] Based on the above embodiments, as an optional embodiment, the target map propagation matrix is ​​determined based on the spatial proximity matrix, road network connectivity matrix, functional similarity matrix, pedestrian flow correlation matrix, electrical topology matrix, and irradiance correlation matrix, including:

[0099] The fusion adjacency matrix is ​​determined by weighted calculation based on the spatial proximity matrix, road network connectivity matrix, functional similarity matrix, pedestrian flow correlation matrix, electrical topology matrix, and irradiance correlation matrix, as well as the corresponding preset learnable weight parameter group.

[0100] Based on two pre-set sets of learnable node embedding parameters, an adaptive adjacency matrix is ​​determined.

[0101] The target graph propagation matrix is ​​determined based on the fused adjacency matrix and the adaptive adjacency matrix.

[0102] Specifically, in the embodiments of this application, six types of adjacency matrix sets are constructed on the same node set. .in: Represents spatial proximity relationships based on the distance between node centroids; It represents the road network connectivity based on the shortest path or traffic impedance on the road map; This represents the functional similarity relationship based on the proportional vector similarity of POIs; Indicates time-based The correlation between OD flow and turnstile intensity in pedestrian flow; This indicates the electrical topology relationship based on the feeder-substation topology distance and the same feeder indication; This represents the irradiance correlation based on the spatial correlation of irradiance / cloud motion at time t.

[0103] Furthermore, it is also possible to... The unified implementation of sparsity reduction and normalization involves the following steps:

[0104] 1) k-nearest neighbor / Top-K or threshold truncation; 2) Row-wise normalization; 3) Timestamp alignment: ensure that the two types of time-varying matrices are consistent with the "node × time" index; 4) Dimensional consistency check: check with the number of nodes. Consistent.

[0105] Furthermore, multi-relation fusion and adaptive adjacency are performed. Fusion adjacency and adaptive adjacency are conducted based on the six types of adjacency matrices mentioned above. To support unified invocation of subsequent models, the interface artifacts can be defined as follows:

[0106] Merge Adjacency Matrix : This represents the convex combination adjacency obtained after assigning learnable weights to each relation in Table 3;

[0107] Adaptive Adjacency Matrix : Represents similarity adjacency obtained by automatically learning parameters from node embeddings;

[0108] Target graph propagation matrix According to a fixed coefficient Fusion and .

[0109] More specifically, learnable weight parameters that are non-negative and sum to 1 are assigned to each relation, and a weighted fusion graph, i.e., a fusion adjacency matrix, is calculated. It can be represented as:

[0110] ;

[0111] in, .

[0112] Simultaneously, two pre-defined sets of learnable node embedding parameters are introduced. , ,in, Two sets of learnable node embedding parameters are specifically used to automatically learn a similarity adjacency map from the data, where each row represents a city region node. d 3D vector. Adaptive adjacency is obtained through node embedding learning. Therefore, it can be calculated using a proportional coefficient. Forming the target graph propagation matrix Finally, we can obtain the node set, node static attributes, and propagation matrix. The photovoltaic-enhanced city map structure is used for subsequent spatiotemporal modeling.

[0113] Preferably, the regional nodes adopt equal-area hexagonal grids (e.g., the grid side length can be selected in the range of 250m to 1000m). Use Gaussian kernel and row Nearest neighbor sparsification; Calculate and add turning penalties at intersections on the roadside-weighted graph; It is obtained from the correlation coefficient of the irradiation time window or its kernel function.

[0114] In one embodiment of this application, a city center area can be selected as the test area. The space is discretized using equal-area hexagonal statistical units with a side length of approximately 500m, forming a 6×7 grid with N=42 regional nodes as the node set. Data including road centerlines, points of interest (POIs), public transport gates / origins (ODs), meteorological and solar irradiance data, power distribution feeder-substation topology, and historical load data are clipped to the test area under a unified plane coordinate system and aggregated using a 30-minute time granularity and node index. Static attributes (building density, road centrality, POI principal components, resident population density, photovoltaic potential, etc.) are generated and stored according to Table 2. Six types of relationship matrices are calculated on the same node set and uniformly sparsity and normalization are applied. The six types of relationships are then plotted with different line types to obtain the final result. Figures 2 to 4 The city map shown is a multi-relationship map. Figures 2 to 4 This document showcases the multi-relationship city map based on six types of adjacency matrices, illustrated in three attached diagrams for clarity. Figure 2 The graphs show the spatial proximity matrix and the pedestrian flow relationship matrix. Figure 3 The graphical representations of the road network connectivity matrix and the electrical topology matrix are shown. Figure 4 The graphs of the functional similarity matrix and the irradiance correlation matrix are displayed. The colors of the lines (red, yellow, and blue) represent the weight of the association between nodes, i.e., strong, medium, and weak levels, respectively; connecting lines indicate that there is an association.

[0115] Furthermore, taking the above six types of relationships as input, a graph-side interface is formed following a fusion-adaptation-propagation process: first, learnable weights are assigned to each relationship. ,through Obtain the merged adjacency, and then embed the nodes. and calculate Finally press The propagation matrix is ​​obtained by linear combination. This data is used by the subsequent spatiotemporal graph neural network ST-GNN, and the output includes six types of adjacency matrix files and optional matrices. , , And a time-synchronized index file, the structure and naming of which follow Table 4.

[0116] It should be noted that regarding the output results, it is necessary to check that the dimensions of each matrix are consistent, the row normalization is valid, and the timestamps of the time-varying matrices match the feature indices; each node in At least a Top-K in-degree is required to avoid isolated nodes; differential privacy noise can be added to node-level load / network power before exporting; static attributes do not contain personal identifiers. By defining a five-layer object structure—data, node, relationship, interface, and output—the fields and update criteria are clearly defined.

[0117] The method in this application embodiment simultaneously constructs six types of relationships on the same node set: spatial proximity, road network connectivity, functional similarity, pedestrian flow association, electrical topology, and irradiance correlation. A propagation matrix is ​​formed through a convex combination of learnable weights and adaptive adjacency. It can realize multi-mechanism relationships and adaptive propagation, covering explicit and implicit coupling. Explicit relationships ensure interpretable physical / urban mechanisms, and adaptive adjacency automatically fills in unobserved or time-varying cross-regional connections (such as synchronous fluctuations under cloud belts and resonance brought by PV penetration through co-feed lines), significantly enhancing the model's generalization and robustness to complex urban scenarios and abnormal days (holidays, event days, extreme weather). For newly connected PV areas, it can also complete cold start initialization by leveraging multi-relationship similarity, shortening the online convergence time.

[0118] Furthermore, in the embodiments of this application, step S3 involves constructing target temporal feature information and supervising target implementation. Specifically, the nearest nodes in each region can be... Historical power load, weather and irradiance (including lagged and rolling statistics), holiday indicators, time-of-use pricing, pedestrian flow intensity index, and PV static attributes (installed density, roof suitability, component orientation / tilt angle, etc.) are stitched together into a tensor. This yields the target's time-series characteristics. At this point, the monitoring target is the net power load. , This represents the total power load of the regional nodes. This represents the grid-connected photovoltaic power of the regional nodes. Optionally, multi-task joint learning can also be performed to simultaneously predict... and and with Participate in the main loss.

[0119] More specifically, let the length of the history window be... The predicted step size is For each time index Construct the input tensor Among them, feature dimension It consists of the following components, and is standardized dimension by dimension:

[0120] a) Historical sequence: Node-level historical net load (or the historical values ​​of total load and grid connection power respectively), 1 to 3 hysteresis terms may be added;

[0121] b) Meteorology / Irradiance: Temperature, relative humidity, wind speed, precipitation, solar irradiance parameters (GHI / DNI / DHI), cloud cover and cloud motion indices;

[0122] c) Travel and Prices / Calendar: Travel intensity index, holiday indications, time-of-use electricity pricing, event indications (events / performances);

[0123] d) Static attribute embedding: The static attributes in Table 2 are linearly mapped or embedded and then concatenated to each time step;

[0124] e) Time encoding: Sine and cosine encoding of positions within hours / weeks. The supervision objective is: That is, the future The node-level net load of the step. Among them, the prediction step size... .

[0125] In the embodiments of this application, before executing step S4, it is necessary to pre-train the spatiotemporal graph neural network using the prepared target training data samples and corresponding power load data labels until a trained spatiotemporal graph neural network is obtained, thus obtaining the power load prediction model. The target training data samples include target graph propagation matrix samples and target time-series feature information samples.

[0126] Specifically, in the embodiments of this application, model training can be performed using a spatiotemporal joint modeling and multi-step prediction approach. A spatiotemporal graph neural network (ST-GNN) is employed, using a temporal convolutional network (TCN) for temporal modeling and a diffusion graph convolution for spatial propagation, to process the target temporal feature information and the target graph propagation matrix. Perform forward computation and output the future. The net load at the step node level and its administrative region / citywide aggregation can be performed; key peak periods (such as 7 to 10 am and 6 to 9 pm) can be weighted in the loss.

[0127] Based on the above embodiments, as an optional embodiment, the loss function used by the power load forecasting model during model training is:

[0128] ;

[0129] In the formula, This represents the power load forecast information for all regional nodes; This represents the actual power load information for all regional nodes; This indicates the spatial proximity information between regional nodes; This indicates the differences in net power load between regional nodes; These represent preset coefficients; This represents the mean absolute error function; This represents the root mean square error function.

[0130] Specifically, in the embodiments of this application, during the model training process, the overall structure of the power load forecasting model adopts a two-layer network structure of stacked spatiotemporal blocks (ST-Blocks) coupled with a "time module + spatial module", which specifically includes:

[0131] a) Temporal Module (TCN): One-dimensional causal convolution, with a kernel size of up to 3, an inflation coefficient sequence of 1 and 2, and includes residuals and layer normalization;

[0132] b) Spatial Module (Diffusion Graph Convolutional DGCN): Spreads the matrix samples in the target graph. Up The number of channels in a diffusing convolution can range from 32 to 64.

[0133] c) Gating and residuals: The time / space output is gated and fused by GLU or Sigmoid, and then added to the block input as a residual;

[0134] d) Readout layer (parallel decoding): uses 1×1 convolution to directly output Multi-channel steps are used to generate prediction results. .

[0135] Furthermore, the loss calculation during model training can employ multi-target loss with weighting for typical peak segments; that is, the loss function can be expressed as:

[0136] ;

[0137] The third term is a spatial smoothing regularization term, which is used to constrain the differences in net load forecasts between adjacent or strongly correlated nodes.

[0138] In one alternative implementation, to improve the prediction accuracy of typical peak periods such as the morning rush hour (e.g., 7–10 a.m.) and the evening rush hour (e.g., 6–9 p.m.) on weekdays, it is possible to... and The calculation introduces a time weighting coefficient. Specifically, when time step When it falls within the above-mentioned peak segment, take Other time steps At this point, the error can be weighted and calculated using the following formula:

[0139] ;

[0140] ;

[0141] here, This is an optional implementation for time-weighted calculations during typical peak hours such as morning and evening rush hours. Within the peak period... Take at other times , All can be determined by the validation set.

[0142] Specific model training parameters include: dividing all target training data samples into training, validation, and test sets according to time; where standardization uses only training segment statistics, and a sliding window is used to form the samples. The optimizer is Adam (cosine annealing or adaptive decay based on the validation set), and the batch size can range from 32 to 128; the number of training epochs is 20 to 100, and early stopping is performed on the validation set.

[0143] Further, perform inference and aggregate the output. During online inference, at regular intervals... Get the latest from the cache Step input and current Generate the future Step-level prediction results The prediction results can be derived at the following three levels:

[0144] Node level: Net load curve for each statistical unit; Administrative region aggregation: Summation / weighted average by regional index; Citywide total: Summation of all nodes and provision of interval estimates (optional 50 / 90 quantile). The output format is consistent with Table 4, and the interface is published in both API and report formats.

[0145] Table 4

[0146]

[0147] In one specific embodiment of this application, the typical data from the foregoing embodiments is still used, and the output result is as follows: Figure 5 and Figure 6 As shown, Figure 5 This is a schematic diagram illustrating the predicted net electricity load for the city. Figure 6 A schematic diagram illustrating the predicted output of urban photovoltaic power. As can be seen from the diagram, the model prediction results provided in this application embodiment are very close to the actual data, demonstrating high testing accuracy.

[0148] Simultaneously, cold start and missing test handling can be performed during model training. Specifically, for new nodes or short historical nodes, similarity can be calculated based on static attributes and functional similarity. Row vectors retrieve the most similar ones from the training set. Each node is initialized with its embedding and temporal module state using similarity weighting; fine-tuning is performed with small samples after obtaining the first week's observations. Input features with short-term missing data are recovered using similar daily data analysis and a three-point interpolation algorithm.

[0149] The method in this application realizes multi-step prediction and critical peak segment protection, directly serving scheduling and energy storage: the model natively supports multi-step parallel / autoregressive decoding, and can introduce loss weighting for critical peak segments such as 7-10 am and 18-21 am. The output includes both node-level curves and aggregated results for administrative regions / the whole city, meeting the timeliness requirements of power supply zone rolling plans, virtual power plant aggregation control, and energy storage charging and discharging strategy optimization; on the operation side, stable multi-step trajectories can be used to reduce temporary start-stop and unnecessary backup, which can effectively alleviate the peak shaving pressure caused by "midday digging - evening climbing".

[0150] Furthermore, in the embodiments of this application, in step S4, the target graph propagation matrix and target time series feature information are input into the power load prediction model. After the power load prediction model identifies and predicts, the power load prediction results of the city to be tested can be obtained quickly.

[0151] The urban power load forecasting method of this application divides the urban area map into nodes and utilizes Euclidean distance, electrical topology information, shortest path length, functional similarity, traffic flow, and solar irradiance information between regional nodes to simultaneously construct multiple types of multi-source adjacency relationships on the same regional node set. These relationships can include spatial proximity, electrical topology, road network connectivity, functional similarity, pedestrian flow correlation, and irradiance correlation. This data is then combined with time-series feature data constructed from multi-source observations that change over time to form multi-source spatiotemporal map data. This data is then used by the power load forecasting model trained by the spatiotemporal map neural network for identification and prediction. This significantly enhances the model's generalization ability and robustness in predicting power load in complex urban scenarios and on abnormal days (such as holidays, event days, and extreme weather days), greatly improving the accuracy and precision of urban power load forecasting.

[0152] Based on the above embodiments, as an optional embodiment, the spatial proximity matrix of all associated regional nodes is determined based on the Euclidean distance between each regional node, including:

[0153] The Gaussian kernel weight matrix is ​​obtained by calculating the Euclidean distance between nodes in each region using the Gaussian kernel weight function.

[0154] Sparsification is performed on each row of the Gaussian kernel weight matrix to obtain the spatial proximity matrix.

[0155] Specifically, in the embodiments of this application, a spatial proximity matrix is ​​determined. The specific implementation method can be as follows:

[0156] Calculate the Euclidean distance between the centroids of nodes in the region. And substitute it into the Gaussian kernel weight function, and take the Gaussian kernel weight. ,in The length scale, from which we can obtain the length scale from each The Gaussian kernel weight matrix is ​​formed.

[0157] Furthermore, perform the following on the Gaussian kernel weight matrix: Nearest neighbor sparsification, that is, retaining the previous row of each row in the matrix. The maximum weight, The range of values ​​can be set to At the same time, those less than the threshold The elements are set to zero, and then normalization is performed row by row to obtain the spatial proximity matrix. .

[0158] The method in this application embodiment, by combining Gaussian kernel weight function calculation and matrix sparsification, can optimize spatial analysis efficiency. It uses Gaussian kernel to enhance the expressive power of the weight matrix and sparsification to reduce computational complexity. The synergy of the two can improve the efficiency and accuracy of large-scale spatial data processing.

[0159] In the embodiments of this application, the road network connectivity neighbor matrix is ​​calculated. The specific implementation method can be as follows:

[0160] A weighted graph with road centerlines as edges is constructed on the road network of the city under test, and Dijkstra's algorithm is used to calculate the shortest path length between regional nodes. The base weight of each edge is taken as the geometric length of the road's centerline. Based on this, a constant penalty term is added to each turning action at intersections along the path. Defined as:

[0161] ;

[0162] in, For the edge geometric length, For the node To the node The number of turns in the shortest path. Then, the shortest path length is mapped to road network association weights:

[0163]

[0164] in This is a scale parameter used to adjust the sensitivity of distance to weight decay.

[0165] Based on the above embodiments, as an optional embodiment, a functional similarity matrix is ​​determined based on the functional similarity between each regional node, including:

[0166] Based on the POI data of each regional node, a multidimensional POI ratio vector for each regional node is determined; the POI data is determined based on the number of points of interest in each social function category possessed by each regional node.

[0167] The functional similarity matrix is ​​determined by calculating the similarity between pairs of nodes using the multidimensional POI proportional vector of each region node.

[0168] Specifically, in the embodiments of this application, a functional similarity matrix is ​​determined. The specific implementation method can be as follows:

[0169] Construct a POI scale vector for each node using the POI data of each region node. The POI data specifically targets seven social function categories: residential, commercial, industrial, educational, healthcare, transportation hubs, and public administration. The POI ratio vector... This involves mapping the original Points of Interest (POI) labels from the city map to this... K kind( K =7), record region nodes i POI tags in the first k The class count is This process can be represented as:

[0170] ;

[0171] ;

[0172] In the formula, K Total number of function types; This is a smoothing coefficient used for Laplacian smoothing when the count for certain categories is 0, avoiding excessively sparse or zero-valued proportions. It is used to smooth the data for each region node. By concatenating the rows, you can obtain the functional structure matrix. .

[0173] Furthermore, to avoid instability caused by all zeros or sparsity, smoothing and normalization are performed. First, for each proportional vector... The data underwent L1 normalization and dimensionality reduction using Principal Component Analysis (PCA) to reduce the data to a minimum. Dimensionality. Then, the cosine similarity between the proportional vectors after dimensionality reduction of each region node is calculated, that is:

[0174] ;

[0175] Finally, after preserving Top-K similarity and performing normalization, the functional similarity matrix is ​​obtained. .

[0176] The method in this application collects the functional category similarity of points of interest (POIs) related to power load activities on each regional node, constructs POI vectors and functional similarity matrices, and captures the inherent correlation between functional electricity consumption in different spatial areas of the city, which is beneficial to further improve the accuracy and robustness of urban power load forecasting.

[0177] In the embodiments of this application, a pedestrian flow correlation matrix is ​​determined. The specific implementation method can be as follows:

[0178] With time steps consistent with power load OD flow or gate strength ,structure .

[0179] If the data is unidirectional, then symmetry is used, i.e. Noisy data is smoothed using a 3-point moving average, thus allowing the data to be processed... Constructing a population flow correlation matrix .

[0180] In the embodiments of this application, a pedestrian flow correlation matrix is ​​constructed by collecting OD flow or gate intensity at each regional node, capturing the inherent correlation between electricity consumption for transportation in different spatial areas of the city, which is beneficial to further improve the accuracy of subsequent urban power load forecasting.

[0181] Based on the above embodiments, as an optional embodiment, determining the electrical topology matrix associated with all regional nodes based on the electrical topology information between each regional node includes:

[0182] Based on the electrical topology information between each area node, determine the shortest electrical path between each area node and the indication information of the same feeder;

[0183] The shortest electrical path between each regional node and the indication information of the same feeder are mapped to a weight matrix between nodes to obtain the electrical topology matrix.

[0184] Specifically, in the embodiments of this application, the electrical topology matrix is ​​determined. The specific implementation method can be as follows:

[0185] Based on the feeder-substation topology diagram of the city under test, calculate the shortest electrical path between the nodes in the two areas, i.e., the electrical line topology distance. and feeder indication information It is understandable that if the co-feedline indicator between the two regional nodes... =0 indicates that the feeder is not conducting; if the feeder indicator between the two area nodes is 0, it means that the feeder is not conducting. =1 indicates that the feeder is in a conductive state.

[0186] Furthermore, take ,in It is used to control the attenuation of electrical signals across the feeder.

[0187] Optionally, if the regional nodes only have the mapping from distribution substations to substations, the node affiliation can be determined by the nearest feeder / substation before calculation.

[0188] The method in the embodiments of this application constructs an electrical topology matrix by collecting the shortest electrical path and co-feeder indication information between each regional node, thereby capturing the inherent correlation of power consumption in the power distribution network of different spatial areas of the city, which is beneficial to further improve the accuracy of subsequent urban power load forecasting.

[0189] In the embodiments of this application, the irradiation correlation matrix is ​​calculated. The specific implementation method can be as follows:

[0190] Utilize each region node in the sliding window The correlation coefficient between nodes in the solar irradiance parameter sequence within the range is calculated. Specifically, solar irradiance parameters include global horizontal irradiance (GHI), direct normal irradiance (DNI), and horizontal diffuse irradiance (DHI). To preset the time step, Hour.

[0191] It should be noted here that if the cloud motion vector can be obtained... If so, then the time lag caused by the movement of the cloud system along this direction to the irradiation sequence can be considered.

[0192] More specifically, for regional nodes With nodes Let their two-dimensional plane coordinates be respectively , Projecting the displacement vectors between nodes onto the direction of cloud movement, the "projection distance" can be defined as:

[0193]

[0194] When node Located at node Downstream of cloud movement (i.e. When the cloud system reaches the node Compared to nodes This will result in a time lag. Based on this, the advection correction time shift can be defined as:

[0195] ;

[0196] in, Indicates cloud system from node Translate to node The required time. The irradiation sequence of downstream nodes will be arranged according to... By performing time shifting (e.g., through interpolation), irradiance fluctuations at different nodes can be "aligned." After completing this advection correction, further adjustments can be made within the time window. Calculate the Pearson correlation coefficient between the two-node irradiation sequences. To construct an irradiance correlation matrix to characterize changes in irradiance synchronicity caused by cloud movement.

[0197] Furthermore, define .in, for Pearson correlation coefficient of radiation sequences within the window; For temperature coefficient, Finally, based on Construct an initial matrix, retain the Top-K correlation of the matrix, and normalize each row to obtain the irradiation correlation matrix. .

[0198] The method in this embodiment of the application achieves precise modeling of photoelectric scenes and direct net load: it introduces irradiance correlations onto a unified city map structure. This approach combines the static PV attributes in subsequent time-series features to transform the chain effect of "cloud system-irradiance-PV output-net load" into a learnable graph propagation. The monitoring target can be directly monitored using net load, or a consistent net load constraint can be obtained through multi-task collaboration of total load / PV output, thereby maintaining predictive stability in photoelectric sensitive scenarios such as sunny / cloudy transitions, cloud gap effects, and rapid shading. In engineering, there is no need for household-by-household measurements; it can be implemented solely based on node-level aggregation and publicly / quasi-public irradiance and ledger data.

[0199] In one specific embodiment of this application, an urban power load forecasting system is also provided, which can adopt a layered architecture of "data layer-layer-model layer-service layer" and be deployed in a data center or cloud platform. The inputs are the generated photovoltaic-enhanced urban map structure (i.e., the target map propagation matrix) and multi-source time-series / static data (i.e., target time-series feature information), and the outputs are multi-step net power load forecasting and interpretation information at the node level, administrative region aggregation level, and city-wide total.

[0200] More specifically, the system may include a data acquisition and standardization module: connecting to roads, buildings, POIs, gate / OD flows, weather and solar irradiance, time-of-use electricity pricing / calendars, historical load and grid power, etc.; completing coordinate unification, time alignment, missing and anomaly repair, and code anonymization; forming a "node × time" master table and a static attribute table. Time-series data can be updated incrementally in 5-60 minute increments, and static data can be updated weekly / monthly.

[0201] It also includes a node mapping and static attribute module: discretizing the city into fixed-type statistical units (one of the following: equal-area hexagons / street communities / functional areas / distribution area aggregations), generating node IDs, and calculating and maintaining static attributes such as building density, road centrality, POI principal components, population, and photovoltaic potential.

[0202] It also includes a multi-relational adjacency building block: computed and stored on the same set of nodes. .in and Updates are performed on a minute-by-minute basis, while other relationships are updated on a daily / weekly basis; relationships are persisted in a graph database / object storage as sparse edge sets.

[0203] It also includes a fusion and adaptation module: as described in the aforementioned embodiments, learnable weight parameters are assigned to each relation matrix. ,through Determine Calculate the fused adjacency matrix And an adaptive adjacency matrix is ​​generated based on node embedding. ,according to Output the target graph propagation matrix.

[0204] It also includes a feature tensor and model service module: constructed based on a sliding window. (Including historical net / total load and grid power, weather / irradiance, travel, price / calendar, time encoding and static embedding), using the ST-Block (a combination of TCN and diffusion graph convolution) model for training and online inference; supports parallel decoding output of future... The system is predicted to use an interface in the form of REST / gRPC, with latency meeting the requirements for online applications in the second range.

[0205] It also includes a publishing and interpretation module: publishing forecasts via API / reports / dashboards; providing visualization of relationship weights, node contribution and peak error profiles; adding differential privacy noise to node-level results as needed and implementing RBAC access control and auditing.

[0206] It also includes a monitoring and drift update module: continuously monitoring MAE / RMSE / MAPE, peak MAE, inference latency, and drift statistics. ;when At that time, rolling retraining is triggered, and the time module and readout layer are quickly updated; full updates of graph-containing convolutions and layers are performed periodically. .

[0207] The system in this application embodiment achieves online adaptive and drift management, ensuring long-term availability: through built-in drift detectors such as CUSUM / distribution distance, it supports a graded update strategy from "lightweight fine-tuning" to "periodic full retraining" to address concept drift caused by seasonal changes, cloud seasonality, PV installation growth, and electricity pricing policy adjustments; spatial smoothing / sparse regularization is added to the training objective to suppress overfitting, ensuring the model remains robust under conditions of missing data, temporary anomalies, and changes in caliber; combined with node-level aggregation and differential privacy (optional), it ensures both controllable long-term operation and maintenance costs and data compliance. Meanwhile, this system is interpretable and engineering-friendly, easy to audit and promote: the model outputs a heatmap of relationship weights and regional contribution decomposition, which can clearly answer "which type of relationship dominates the net load change in which area" (such as radiation dominance caused by the passage of cloud belts, or electrical dominance caused by PV penetration through the same feeder), which facilitates the formation of auditable operation and control basis; the system adopts a modular design (data standardization → graph construction → fusion → ST-GNN → release), supports batch / streaming inference and center-edge collaborative deployment, can be quickly launched at the city level without modifying the existing metering infrastructure, and can be replicated and promoted in multiple cities according to the same framework.

[0208] The urban power load forecasting device provided in this application is described below. The urban power load forecasting device described below can be referred to in correspondence with the urban power load forecasting method described above.

[0209] Figure 7 This is a schematic diagram of the urban power load forecasting device provided in the embodiments of this application, as shown below. Figure 7 As shown, it includes:

[0210] The partitioning module 10 is used to divide the regional map of the city to be tested into a regional map structure centered on multiple regional nodes according to a preset regional range.

[0211] The first processing module 20 is used to determine the target graph propagation matrix based on the Euclidean distance between each regional node, electrical topology information, and other multi-source association information; the other multi-source association information includes at least one of the following: shortest path length, functional similarity, traffic flow, and irradiance information;

[0212] The second processing module 30 is used to determine the target time-series characteristic information based on the power load data, meteorological and radiation data, comprehensive electricity consumption information of personnel and photovoltaic power generation installed capacity of each regional node in the most recent historical period.

[0213] Prediction module 40 is used to input the target map propagation matrix and target time series feature information into the power load prediction model to obtain the power load prediction results of the city to be tested output by the power load prediction model;

[0214] The power load forecasting model is trained based on the target training data samples and the corresponding power load data labels; the target training data samples include target graph propagation matrix samples and target time series feature information samples.

[0215] It is understood that the detailed functional implementation of each of the above units / modules can be found in the description in the aforementioned method embodiments, and will not be repeated here.

[0216] It should be understood that the above-described device is used to execute the methods in the above embodiments. The implementation principle and technical effect of the corresponding program modules in the device are similar to those described in the above methods. The working process of the device can be referred to the corresponding process in the above methods, and will not be repeated here.

[0217] The urban power load forecasting device of this application, through the division of urban area map nodes, utilizes Euclidean distance, electrical topology information, shortest path length, functional similarity, traffic flow, and solar irradiance information between area nodes to simultaneously construct multiple types of multi-source adjacency relationships on the same area node set. These relationships can include spatial proximity, electrical topology, road network connectivity, functional similarity, pedestrian flow correlation, and irradiance correlation. This data is then combined with time-series feature data constructed from multi-source observations that change over time to form multi-source spatiotemporal map data. This data is then used by the power load forecasting model trained by the spatiotemporal map neural network for identification and prediction. This significantly enhances the model's generalization ability and robustness in predicting power load under complex urban scenarios and abnormal days (such as holidays, event days, and extreme weather days), greatly improving the accuracy and precision of urban power load forecasting.

[0218] Based on the methods in the above embodiments, this application provides an electronic device, such as... Figure 8 As shown, the electronic device may include a processor 810, a communications interface 820, a memory 830, and a communication bus 840, wherein the processor 810, the communications interface 820, and the memory 830 communicate with each other via the communication bus 840. The processor 810 can call logical instructions in the memory 830 to execute the methods in the above embodiments.

[0219] Furthermore, the logical instructions in the aforementioned memory 830 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application.

[0220] Based on the methods in the above embodiments, this application provides a computer-readable storage medium storing a computer program that, when run on a processor, causes the processor to execute the methods in the above embodiments.

[0221] Based on the methods in the above embodiments, this application provides a computer program product that, when run on a processor, causes the processor to execute the methods in the above embodiments.

[0222] It is understood that the processor in the embodiments of this application can 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, transistor logic devices, hardware components, or any combination thereof. A general-purpose processor can be a microprocessor or any conventional processor.

[0223] The method steps in this application embodiment can be implemented in hardware or by a processor executing software instructions. The software instructions can consist of corresponding software modules, which can be stored in random access memory (RAM), flash memory, read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), registers, hard disks, portable hard disks, CD-ROMs, or any other form of storage medium known in the art. An exemplary storage medium is coupled to the processor, enabling the processor to read information from and write information to the storage medium. Of course, the storage medium can also be a component of the processor. The processor and the storage medium can reside in an ASIC.

[0224] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially as a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted through the computer-readable storage medium. The computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state disk (SSD)).

[0225] It is understood that the various numerical designations used in the embodiments of this application are merely for the convenience of description and are not intended to limit the scope of the embodiments of this application.

[0226] It should be understood that expressions such as “comprising” and “may include” used in this application indicate the existence of the disclosed functions, operations, or constituent elements, and do not limit one or more additional functions, operations, and constituent elements. In this application, terms such as “comprising” and / or “having” are to be interpreted as indicating a particular characteristic, number, operation, constituent element, component, or combination thereof, but not to exclude the existence or possibility of adding one or more other characteristics, numbers, operations, constituent elements, components, or combinations thereof.

[0227] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for predicting urban electricity load, characterized in that, include: The regional map of the city to be tested is divided into a regional map structure centered on multiple regional nodes according to the preset regional range. The target graph propagation matrix is ​​determined based on the Euclidean distance between each of the regional nodes, electrical topology information, and other multi-source association information; the other multi-source association information includes at least one of the following: shortest path length, functional similarity, traffic flow, and irradiance information. Based on the power load data, meteorological and radiation data, comprehensive electricity consumption information of personnel, and photovoltaic power generation installed capacity density of each regional node in the most recent historical period, target time-series characteristic information is determined. The target graph propagation matrix and the target time series feature information are input into the power load prediction model to obtain the power load prediction result of the city to be tested output by the power load prediction model. The power load prediction model is trained based on target training data samples and corresponding power load data labels; the target training data samples include target graph propagation matrix samples and target time-series feature information samples. The other multi-source association information includes shortest path length, functional similarity, traffic flow, and irradiance information; the determination of the target graph propagation matrix based on the Euclidean distance between each of the regional nodes, electrical topology information, and other multi-source association information includes: Based on the Euclidean distance between each of the said region nodes, determine the spatial proximity matrix that associates all said region nodes; Based on the electrical topology information between each of the said regional nodes, an electrical topology matrix relating all said regional nodes is determined; Based on the shortest path length between each of the said regional nodes, determine the road network connectivity matrix that associates all said regional nodes; Based on the functional similarity between each of the said regional nodes, a functional similarity matrix is ​​determined that associates all said regional nodes. Based on the traffic flow between each of the said regional nodes, a passenger flow association matrix is ​​determined that relates all said regional nodes; Based on the irradiation information of each of the said regional nodes, an irradiation correlation matrix relating all said regional nodes is determined; The target map propagation matrix is ​​determined based on the spatial proximity matrix, the road network connectivity matrix, the functional similarity matrix, the pedestrian flow correlation matrix, the electrical topology matrix, and the irradiance correlation matrix.

2. The urban power load forecasting method according to claim 1, characterized in that, The determination of the target map propagation matrix based on the spatial proximity matrix, the road network connectivity matrix, the functional similarity matrix, the pedestrian flow correlation matrix, the electrical topology matrix, and the irradiance correlation matrix includes: The fusion adjacency matrix is ​​determined by weighting the spatial proximity matrix, the road network connectivity matrix, the functional similarity matrix, the pedestrian flow correlation matrix, the electrical topology matrix, and the irradiance correlation matrix, as well as the corresponding preset learnable weight parameter group. Based on two pre-set sets of learnable node embedding parameters, an adaptive adjacency matrix is ​​determined. The target graph propagation matrix is ​​determined based on the fused adjacency matrix and the adaptive adjacency matrix.

3. The urban power load forecasting method according to claim 1, characterized in that, Determining the spatial proximity matrix of all associated regional nodes based on the Euclidean distance between each of the regional nodes includes: The Gaussian kernel weight matrix is ​​obtained by calculating the Euclidean distance between each node in the region using a Gaussian kernel weight function. The spatial proximity matrix is ​​obtained by sparsifying the elements of each row in the Gaussian kernel weight matrix.

4. The urban power load forecasting method according to claim 1, characterized in that, Determining the functional similarity matrix associated with all the regional nodes based on the functional similarity between each of the regional nodes includes: Based on the POI data of each of the aforementioned regional nodes, a multidimensional POI ratio vector for each of the aforementioned regional nodes is determined; the POI data is determined based on the number of points of interest in each of the aforementioned regional nodes across social function categories. The functional similarity matrix is ​​determined by calculating the similarity between pairs of nodes using the multidimensional POI ratio vector of each region node.

5. The urban power load forecasting method according to claim 1, characterized in that, The step of determining the electrical topology matrix associated with all the regional nodes based on the electrical topology information between each of the regional nodes includes: Based on the electrical topology information between each of the regional nodes, determine the shortest electrical path and indication information of the same feeder between each of the regional nodes; The electrical topology matrix is ​​obtained by mapping the shortest electrical path and the indication information of the same feeder between each of the regional nodes into a weight matrix between the nodes.

6. The urban power load forecasting method according to any one of claims 1-5, characterized in that, The loss function used in the training process of the power load forecasting model is: ； In the formula, This represents the power load forecast information for all regional nodes; This represents the actual power load information for all regional nodes; This indicates the spatial proximity information between regional nodes; This indicates the differences in net power load between regional nodes; These represent preset coefficients; This represents the mean absolute error function; This represents the root mean square error function.

7. A city power load forecasting device, characterized in that, include: The partitioning module is used to divide the regional map of the city to be tested into a regional map structure centered on multiple regional nodes according to a preset regional range. The first processing module is used to determine the target graph propagation matrix based on the Euclidean distance between each of the region nodes, electrical topology information, and other multi-source association information; The other multi-source association information includes at least one of the following: shortest path length, functional similarity, traffic flow, and irradiance information; The second processing module is used to determine target time-series characteristic information based on the power load data, meteorological and radiation data, comprehensive electricity consumption information of personnel and photovoltaic power generation installed capacity of each regional node in the most recent historical period. The prediction module is used to input the target graph propagation matrix and the target time series feature information into the power load prediction model to obtain the power load prediction result of the city to be tested output by the power load prediction model; The power load prediction model is trained based on target training data samples and corresponding power load data labels; the target training data samples include target graph propagation matrix samples and target time-series feature information samples. The other multi-source association information includes shortest path length, functional similarity, traffic flow, and irradiance information; the determination of the target graph propagation matrix based on the Euclidean distance between each of the regional nodes, electrical topology information, and other multi-source association information includes: Based on the Euclidean distance between each of the said region nodes, determine the spatial proximity matrix that associates all said region nodes; Based on the electrical topology information between each of the said regional nodes, an electrical topology matrix relating all said regional nodes is determined; Based on the shortest path length between each of the said regional nodes, determine the road network connectivity matrix that associates all said regional nodes; Based on the functional similarity between each of the said regional nodes, a functional similarity matrix is ​​determined that associates all said regional nodes. Based on the traffic flow between each of the said regional nodes, a passenger flow association matrix is ​​determined that relates all said regional nodes; Based on the irradiation information of each of the said regional nodes, an irradiation correlation matrix relating all said regional nodes is determined; The target map propagation matrix is ​​determined based on the spatial proximity matrix, the road network connectivity matrix, the functional similarity matrix, the pedestrian flow correlation matrix, the electrical topology matrix, and the irradiance correlation matrix.

8. An electronic device, characterized in that, Includes memory and one or more processors; The memory is coupled to the one or more processors, and the memory is used to store computer program code, the computer program code including computer instructions; The one or more processors invoke the computer instructions to cause the electronic device to perform the method as described in any one of claims 1-6.

9. A computer-readable storage medium comprising instructions, characterized in that: When the instructions are executed on an electronic device, the electronic device causes the electronic device to perform the method as described in any one of claims 1-6.

Images