Prediction method considering space-time two-dimensional characteristics of space power load

Abstract

The invention discloses a prediction method considering space-time two-dimensional characteristics of a space power load, and belongs to the technical field of power distribution network planning. According to the method, an electric power geographic information system (GIS) is established, I-class cellular load data is converted into II-class cellular load data by using a space electric power load gridding technology, and then a space load prediction model is established by using GAT, DeepTCN and a multi-head attention mechanism; wherein the GAT and the DeepTCN respectively acquire spatial dimension features and time dimension features of the class-II cellular load, deep fusion of the spatial and temporal features of the class-II cellular load is realized by using a multi-head attention mechanism, and a class-II cellular-based spatial load prediction result is output through a full connection layer; an example analysis result verifies the correctness and effectiveness of the method; according to the method, the problem of excessive dependence of the prediction model on a certain dimension feature can be effectively avoided, the prediction precision is improved, and more accurate data support can be provided for lean planning of an urban power grid.

Description

Technical Field

[0001] This invention belongs to the field of power distribution network planning technology, and in particular relates to a prediction method that considers the spatiotemporal dual-dimensional characteristics of spatial power load. Background Technology

[0002] Spatial load forecasting (SLF) is the prediction of the size and location of future loads within a forecast area. It is the basis for urban power grid planning. The results of SLF can provide important references for the configuration of electrical equipment capacity and the determination of installation locations, thereby making the planning and construction of urban power grids more economical and reliable.

[0003] Most existing spatial load forecasting methods focus on improving data quality or optimizing forecasting models to better uncover the temporal development patterns of cell loads, but lack in-depth and sufficient exploration and analysis of the spatial correlation between cell loads. Therefore, there is an urgent need for a new technical solution to address this problem. Summary of the Invention

[0004] The technical problem to be solved by this invention is to provide a prediction method that considers the spatiotemporal dual-dimensional characteristics of spatial power load to solve the technical problem that current research does not adequately consider the spatial influence between type II cell loads. On this basis, the key spatiotemporal characteristics of type II cell loads are deeply fused to avoid the prediction model from over-reliance on a single dimension feature.

[0005] A prediction method considering the spatiotemporal dual-dimensional characteristics of spatial power load includes the following steps, which are performed sequentially:

[0006] Step 1: Establish a power geographic information system (GIS);

[0007] Step 2: Generate a spatial load spatiotemporal information matrix based on type II cells;

[0008] Step 3: Extract the spatial dimension features of type II cell loads based on the graph attention network (GAT);

[0009] Step 4: Extract the temporal dimension features of type II cell loads based on the DeepTCN deep temporal convolutional network;

[0010] Step 5: Fusion of spatiotemporal characteristics of type II cell load;

[0011] Step 6: Construct a prediction model that considers the spatiotemporal characteristics of spatial power load, and predict the spatiotemporal characteristics of power load.

[0012] The specific method for establishing the power geographic information system (GIS) in step one is as follows:

[0013] ① Collect and organize basic data information within the area to be predicted, and establish a power geographic information system (GIS) using the organized basic data information. The basic data information includes satellite images; street maps, grid structures, geographical locations of power plants and substations, specific routes of transmission lines, power supply range of 10kV feeders, land use type information, and historical load data. The power geographic information system (GIS) is used to establish a registration base map, land use information layer, and type II cell layer for the area to be predicted. It is also used to generate a spatial load spatiotemporal information matrix based on type II cells, extract the spatial dimension features of type II cell loads, extract the temporal dimension features of type II cell loads, perform the fusion of spatiotemporal features of type II cell loads, and predict the spatiotemporal dual-dimensional features of power loads. The grid structure refers to the network formed by the interconnection of power sources, transmission lines, busbars, and transformers.

[0014] ② In the power geographic information system (GIS), firstly, the latitude and longitude values ​​of at least four points in the satellite image or street map to be registered are determined in advance, which are more than four points at a set distance threshold apart. Then, using satellite image, street map, grid structure, geographical location of power plants and substations, specific route of transmission lines, land use information and historical load data, commercial professional software of geographic information system is used to establish the registration base map of the area to be predicted.

[0015] In the registered base map, the area to be predicted is divided according to the different land use types set, and a land use information layer is established;

[0016] The prediction area is divided according to the power supply range of the 10kV feeder, and a type I cell layer is established. Then, the prediction area is divided according to a regular grid of equal-sized squares with a side length of 300m, and a type II cell layer is established.

[0017] The different land use types mentioned include residential land, commercial land, cultural and entertainment land, industrial land, administrative office land, green land, municipal facilities land, and special land.

[0018] The specific method for generating the spatial load spatiotemporal information matrix based on type II cells in step two is as follows:

[0019] ① The spatial power load gridding technology is used to transform the spatial power load data based on type I cells into spatial power load data based on type II cells at the corresponding time.

[0020] ② Treat each type II cell as a node, the load data of each type II cell as the feature information of the node, and the adjacency relationship between each type II cell as the edge, and generate a spatial load spatiotemporal information matrix based on type II cells. This matrix contains the load values ​​of all type II cells and their corresponding spatial location distribution information.

[0021] The specific method for extracting the spatial dimension features of type II cell loads is as follows:

[0022] ① Generate adjacency matrix A using the spatial load spatiotemporal information matrix based on type II cells;

[0023] Two type II cells are considered as two cell nodes. If they share a common boundary in their spatial distribution, then these two type II cells are considered to be adjacent. This is demonstrated using the adjacency matrix A∈R. N×N The connection relationship between type II cells is represented by R, where R is the set of real numbers, N is the total number of type II cells, and two cell nodes are represented by cell node v. p With cell node v q If we express that the value a in the p-th row and q-th column of the adjacency matrix A is... pq =1,

[0024] Conversely, cell node v p With cell node v q If they are not adjacent, then a pq =0;

[0025] ② Calculate the attention coefficient;

[0026] Calculate the target type II cell node v using the attention mechanism. i With neighboring type II cell node v j The attention coefficient is calculated using the formula shown in formula (1):

[0027] e ij =σ1(α) T [Wh i ||Wh j ]) (1);

[0028] In the formula, e ij For type II cell node v i With neighboring type II cell node v j Attention coefficient; σ1(·) is the Leaky ReLU activation function; α T W is the learnable parameter vector; W is the weight parameter matrix; || denotes vector concatenation; h i h j They are Class II cell nodes v i and v j The characteristics; i and j represent the i-th cell and the j-th cell in all type II cells, respectively;

[0029] ③ Normalize attention weights;

[0030] To make the attention coefficients of different nodes comparable, the softmax function is used to normalize the attention coefficients of different nodes, resulting in the normalized attention coefficient α. ij ;

[0031] ④ Weighted aggregation of neighbor node features;

[0032] To better learn the feature information between each type II cell, a multi-head attention mechanism is applied to perform a weighted summation of the features of neighboring type II cell nodes. The features of each type II cell node are calculated by K independently working attention mechanisms, and the results are concatenated to obtain the final features of the type II cell nodes. The calculation formula is shown in formula (2).

[0033]

[0034] In the formula, h i ′ is a type II cell node v i The features after weighted aggregation; k = 1, 2, ..., K, where K is the number of attention heads; σ²(·) is the sigmoid activation function; j ∈ L i， L i For type II cell node v i The set of all neighboring Class II cells; W represents the normalized attention coefficient for the k-th attention head. k h is the weight parameter matrix for the k-th attention head. j For type II cell node v j Its characteristics.

[0035] The specific method for extracting the time dimension features of type II cell load in step four is as follows:

[0036] ①Calculation of dilated causal convolution;

[0037] The DeepTCN expands the receptive field through the dilation coefficient d to capture the long-term temporal development pattern of type II cell loads; the type II cell load data is temporally unfolded and dilated causal convolution is calculated, as shown in formula (3):

[0038]

[0039] In the formula, y t (l) W represents the type II cell feature vector output by the l-th layer of the network at time t; m = 0, 1, ..., M-1, where M is the kernel size; m (l) The m-th weight parameter matrix of the convolutional kernel in the l-th layer of the network; y t-dm (l-1) denoted as , where is the type II cell feature vector output by the (l-1)th layer of the network at time t-dm; d is the inflation coefficient.

[0040] ② Calculate the time dimension characteristics of type II cell loading;

[0041] The DeepTCN temporal convolutional network extracts the temporal features of type II cellular loads by combining dilated causal convolution and skip connections. A residual block is formed by two dilated causal convolutions. Stacking multiple residual blocks captures the long-term dependencies of type II cellular loads. Then, skip connections linearly superimpose the input data of type II cellular loads with the output features of the residual blocks to obtain the temporal features of type II cellular loads. The calculation formula for skip connections is shown in Equation (4).

[0042] y′ t =y t +x t (4);

[0043] In the formula, y t ' is the time-series feature vector of type II cell load output by the model at time t; y t x is the output eigenvector of the residual block at time t; t Let be the original eigenvector of the type II cell load at time t.

[0044] The specific method for fusing the spatiotemporal features of type II cell loads in step five is as follows:

[0045] A spatiotemporal fusion layer is constructed using a multi-head attention mechanism to deeply extract the spatial and temporal features of type II cellular loads;

[0046] First, linear transformations are performed on the spatial and temporal feature matrices of the type II cell load to generate the Q, K, and V matrices required for each attention head.

[0047] Secondly, for each attention head, the attention weights are obtained by calculating the dot product of matrix Q and matrix K, and then performing scaling and softmax operations.

[0048] Then, the matrix V is weighted and summed using attention weights to obtain the output features of each head;

[0049] Finally, the outputs of all attention heads are concatenated and linearly transformed to obtain the fused spatiotemporal feature matrix of type II cell load.

[0050] In step six, a prediction model considering the spatiotemporal dual-dimensional characteristics of spatial power load is constructed. The specific method for predicting the spatiotemporal dual-dimensional characteristics of power load is as follows:

[0051] A prediction model considering the spatiotemporal dual-dimensional features of spatial power load is established using graph attention network (GAT), deep temporal convolutional network (DeepTCN), and multi-head attention mechanism.

[0052] ① The spatiotemporal matrix of spatial load based on type II cells is used as the input of the prediction model. GAT is used to weight and aggregate the node features of each type II cell with the node features of adjacent cells to extract the spatial dimension features of type II cell load. DeepTCN dilated causal convolution is used to capture long-term dependencies layer by layer. The temporal information of type II cell load is extracted by stacking residual blocks and skip connections to extract the temporal dimension features of type II cell load.

[0053] ② The spatial and temporal features of the type II cell load are weighted and aggregated using the multi-head attention mechanism of the spatiotemporal fusion layer to obtain the spatiotemporal feature matrix of the type II cell load;

[0054] ③ The spatial load prediction results based on type II cells are output through the fully connected layer set by the prediction model.

[0055] Through the above design scheme, the present invention can bring the following beneficial effects:

[0056] This invention is a prediction method considering the spatiotemporal dual-dimensional characteristics of spatial power load. First, it integrates basic data information of the area to be predicted to establish a power geographic information system (GIS). Then, it uses spatial power load gridding technology to transform type I cellular load data into type II cellular load data. Finally, it uses Graph Attention Network (GAT) and Deep Temporal Convolutional Network (DTCN) to predict power load. A spatial load prediction model is established using Convolutional Network (DeepTCN) and a multi-head attention mechanism. GAT and DeepTCN are used to obtain the spatial and temporal features of type II cell loads, respectively. The multi-head attention mechanism achieves deep fusion of the spatiotemporal features of type II cell loads. The spatial load prediction results based on type II cells are output through a fully connected layer. Case studies verify the correctness and effectiveness of the proposed method. The proposed method not only comprehensively considers the spatial and temporal features of type II cell loads but also deeply fuses their key spatiotemporal features, effectively avoiding the problem of over-reliance on a single dimension in the prediction model, thus improving prediction accuracy and providing more precise data support for the lean planning of urban power grids. Attached Figure Description

[0057] The present invention will be further described below with reference to the accompanying drawings and specific embodiments:

[0058] Figure 1 This is a basic principle diagram of the prediction method of the present invention that considers the spatiotemporal dual-dimensional characteristics of spatial power load;

[0059] Figure 2This is a land use information map of the area to be predicted in the prediction method of the present invention, which considers the spatiotemporal dual-dimensional characteristics of spatial power load.

[0060] Figure 3 This is a prediction model diagram in the prediction method of the present invention that considers the spatiotemporal dual-dimensional characteristics of spatial power load;

[0061] Figure 4 This diagram illustrates the extraction process of spatial dimension features of type II cellular load in the prediction method for considering the spatiotemporal dual-dimensional features of spatial power load in this invention.

[0062] Figure 5 The residual block structure diagram of DeepTCN in the prediction method of spatial power load considering the spatiotemporal dual-dimensional characteristics of the present invention;

[0063] Figure 6 This is a comparison chart of the quasi-measured values ​​of type II cell load and the predicted values ​​of various methods in the prediction method of spatial power load considering the spatiotemporal dual-dimensional characteristics of the present invention. Detailed Implementation

[0064] The following uses Figures 1-6 The invention will be further illustrated by examples;

[0065] This invention is a prediction method that considers the spatiotemporal dual-dimensional characteristics of spatial power load. The basic principle diagram is shown below. Figure 1 As shown, the examples of the method include: establishing a power GIS, dataset analysis and processing, constructing a spatial load prediction model, setting error evaluation indicators, error analysis of prediction results, and ablation experiments;

[0066] Based on the prediction method for spatial power load considering the spatiotemporal dual-dimensional characteristics proposed in this invention, historical load data from an administrative region in a city in Northeast my country from 2009 to 2013 were used to predict the Class II cell load value of the region in 2014. To verify the correctness and effectiveness of the proposed method, the prediction results were compared with those of the grey theory method, exponential smoothing method, linear regression method, improved long short-term memory neural network (LSTM) prediction method, graph convolutional network and deep temporal convolutional network combined (GCN-DeepTCN) prediction method, and spatiotemporal graph attention network (STGAT) prediction method. Simultaneously, an ablation experiment was conducted using the dataset of the region to be predicted, decomposing the key modules of the proposed prediction model. By comparing the spatial load prediction accuracy before and after removing these modules, the effectiveness of the key parts of the prediction method was verified. The specific details are as follows:

[0067] 1) Establish a power GIS

[0068] In power geographic information systems (GIS), the first step is to determine the latitude and longitude values ​​of at least four points in the satellite imagery or street map to be registered, which are at least a set distance threshold apart. Then, using satellite imagery, street maps, power grid structures, the geographical locations of power plants and substations, the specific routes of transmission lines, land use information, and historical load data, commercial GIS software, including international mainstream software such as ArcGIS, MapInfo, and QGIS, as well as domestic software such as MapGIS, SuperMap, and GeoScene, is used to establish a registration base map for the area to be predicted.

[0069] Under the registration base map in the power GIS, the areas to be predicted are divided according to eight different land use types: residential land, commercial land, cultural and entertainment land, industrial land, administrative office land, green land, municipal facilities land, and special land, and a land use information layer is established; the areas to be predicted are divided according to the power supply range of 10kV feeders, and a type I cell layer is established; the areas to be predicted are divided according to equal-sized square grids with a side length of 300m, and a type II cell layer is established. Figure 2 This is a land use information map of the area to be predicted;

[0070] 2) Dataset Analysis and Processing

[0071] The sampling interval for each 10kV feeder in the area to be predicted is 5 minutes. First, outlier detection and correction are performed on the historical load data of each type I cell load from 2009 to 2013. Then, the reasonable maximum value of each month in all the corrected type I cell load data from 2009 to 2013 is selected and converted into type II cell load data at the corresponding time using spatial power load gridding technology. Finally, each type II cell is regarded as a node, the load data of each type II cell is used as the feature information of the node, and the adjacency relationship between each type II cell is used as the edge to generate a spatial load spatiotemporal information matrix based on type II cells, thereby obtaining the monthly reasonable maximum value time series of the type II cell spatiotemporal information matrix.

[0072] 3) Construct a spatial load forecasting model

[0073] A prediction model considering the spatiotemporal dual-dimensional characteristics of spatial power load is established using GAT, DeepTCN, and multi-head attention mechanisms. The model diagram is shown below. Figure 3 As shown in the figure, T is the total time step of the type II cell load, and n1 and n2 are the number of GAT layers and the number of residual blocks, respectively.

[0074] First, the spatiotemporal matrix of spatial load based on type II cells is used as the input to the prediction model. GAT (Generative Attraction and Data Type) is then used to weighted aggregate the node features of each type II cell with the node features of its neighboring cells, thereby extracting the spatial dimension features of the type II cell load. The process of extracting the spatial dimension features of the type II cell load is as follows: Figure 4 As shown; secondly, long-term dependencies are captured layer by layer using DeepTCN dilated causal convolution, and temporal information of type II cell loads is extracted by stacking residual blocks and skip connections, thereby realizing the extraction of temporal dimension features of type II cell loads. Figure 5 The diagram shows the residual block structure of DeepTCN. Then, the spatial and temporal features of the type II cell load are weighted and aggregated using the multi-head attention mechanism of the spatiotemporal fusion layer to obtain the spatiotemporal feature matrix of the type II cell load. Finally, the spatial load prediction results based on type II cells are output through a fully connected layer.

[0075] 4) Set error evaluation indicators

[0076] Considering that the error in spatial load forecasting results has a dual characteristic of magnitude and spatial distribution, this invention, in order to evaluate spatial load forecasting methods more rationally and effectively, not only uses the average absolute error and average relative error percentage as two magnitude error evaluation indicators, but also introduces a magnitude-space characteristic error evaluation method. This method cancels out the positive and negative errors in the spatial load forecasting results to obtain the magnitude-space cancellation influence value X. OEIV The remaining uncompensated errors are then summed to obtain the amplitude-space superposition effect value X. REIV The error impact value X is obtained by linearly superimposing the two. EIV As an error evaluation index, X EIV The smaller the value, the less adverse the impact on power grid planning, and the higher the accuracy of the prediction method.

[0077] 5) Error analysis of prediction results

[0078] To make the prediction results more intuitive, the quasi-measured values ​​for the target year, the prediction results of the GCN-LSTM prediction method combining Graph Convolutional Network (GCN) and Long Short-Term Memory Neural Network (LSTM), the GCN-DeepTCN prediction method, the STGAT prediction method, and the prediction results of the method of this invention are displayed in the power GIS. Figure 6 As shown, in power GIS, the size of the cell load value can be represented by the color depth; the darker the color, the larger the load value, and the lighter the color, the smaller the load value.

[0079] Table 1. Evaluation Results of Amplitude Error

[0080]

[0081] The amplitude error evaluation results of various spatial load prediction SLF methods are shown in Table 1. As can be seen from Table 1, compared with the other six prediction methods, the method of this invention has the lowest percentage of both average absolute error and average relative error, meaning that, based on the comparison of amplitude error evaluation indicators, the method of this invention has the highest prediction accuracy. Compared with the GCN-DeepTCN prediction method, the average absolute error and average relative error of the method of this invention are 0.01MW and 2.26% lower, respectively, indicating that GAT is more suitable than GCN for extracting the spatial dimension features of type II cellular loads. Compared with the STGAT prediction method, the average absolute error and average relative error of the method of this invention are 0.027MW and 0.62% lower, respectively. Therefore, it is evident that DeepTCN is superior to GRU neural networks (Gated Recurrent Unit, GRU) when processing time series of type II cellular loads.

[0082] Table 2 Evaluation Results of Spatial Characteristic Errors

[0083]

[0084]

[0085] The amplitude-space characteristic error evaluation results of each SLF method are shown in Table 2. As can be seen from Table 2, the X-axis error of the method of this invention... EIV The accuracy is lower than that of the other six prediction methods, indicating that the method of this invention has the highest prediction accuracy based on the comparison results of the amplitude-space characteristic error evaluation index.

[0086] As can be seen from Tables 1 and 2, the prediction error of the method of the present invention is lower than that of the other six prediction methods, whether it is the amplitude error evaluation index or the amplitude-space characteristic error evaluation index. That is, the prediction effect of the method of the present invention is the best, and it can provide more accurate data support for power grid planning.

[0087] 5) Ablation test

[0088] Ablation experiments were conducted using a dataset of the region to be predicted. The key modules of the prediction model proposed in this invention were decomposed, and the effectiveness of the key parts of the prediction method was verified by comparing the spatial load prediction accuracy before and after removing the module. The prediction results of this invention were compared with those of the GAT prediction method, the DeepTCN prediction method, and the GAT-DeepTCN prediction method (a combination of GAT and DeepTCN). The GAT-DeepTCN prediction method first uses GAT to extract the spatial dimension features of the type II cell load, and then uses DeepTCN to extract the temporal dimension features of the type II cell load for SLF. The ablation experiment results are shown in Tables 3 and 4.

[0089] Table 3 Results of Amplitude Error Ablation Experiment

[0090]

[0091] Table 4 Results of the Spatial Characteristic Error Ablation Experiment

[0092]

[0093] Tables 3 and 4 show that the key components of the model can improve prediction accuracy. Compared with the GAT prediction method, the mean absolute error, mean relative error, and error impact value of the GAT-DeepTCN prediction method are reduced by 0.22MW, 3.27%, and 57.418MW·m, respectively. Compared with the DeepTCN prediction method, the mean absolute error, mean relative error, and error impact value of the GAT-DeepTCN prediction method are reduced by 0.034MW, 3.01%, and 31.204MW·m, respectively. This indicates that comprehensively considering the spatiotemporal characteristics of type II cell loads for SLF results in higher accuracy. Both GAT and DeepTCN are indispensable key components of the model. Compared with the GAT-DeepTCN prediction method, the mean absolute error, mean relative error, and error impact value of the method in this invention are reduced by 0.011MW, 0.4%, and 6.972MW·m, respectively. This demonstrates that introducing a multi-head attention mechanism to deeply fuse the spatiotemporal characteristics of type II cell loads can improve the prediction accuracy of the model.

[0094] The specific embodiments used in this invention have provided a detailed description of the invention, but are not limited to these embodiments. Any obvious modifications made by those skilled in the art based on the teachings of this invention are within the scope of protection of this invention.

Claims

1. A prediction method considering the spatiotemporal dual-dimensional characteristics of spatial power load, characterized in that: Includes the following steps, And the following steps are performed in sequence: Step 1: Establish a power geographic information system (GIS); Step 2: Generate a spatial load spatiotemporal information matrix based on type II cells; Step 3: Extract the spatial dimension features of type II cell loads based on the graph attention network (GAT); Step 4: Extract the temporal dimension features of type II cell loads based on the DeepTCN deep temporal convolutional network; Step 5: Fusion of spatiotemporal characteristics of type II cell load; Step 6: Construct a prediction model that considers the spatiotemporal characteristics of spatial power load, and predict the spatiotemporal characteristics of power load.

2. The prediction method considering the spatiotemporal dual-dimensional characteristics of spatial power load according to claim 1, characterized in that: The specific method for establishing the power geographic information system (GIS) in step one is as follows: ① Collect and organize basic data information within the area to be predicted, and establish a power geographic information system (GIS) using the organized basic data information. The basic data information includes satellite images; street maps, grid structures, geographical locations of power plants and substations, specific routes of transmission lines, power supply range of 10kV feeders, land use type information, and historical load data. The power geographic information system (GIS) is used to establish a registration base map, land use information layer, and type II cell layer for the area to be predicted. It is also used to generate a spatial load spatiotemporal information matrix based on type II cells, extract the spatial dimension features of type II cell loads, extract the temporal dimension features of type II cell loads, perform the fusion of spatiotemporal features of type II cell loads, and predict the spatiotemporal dual-dimensional features of power loads. The grid structure refers to the network formed by the interconnection of power sources, transmission lines, busbars, and transformers. ② In the power geographic information system (GIS), firstly, the latitude and longitude values ​​of at least four points in the satellite image or street map to be registered are determined in advance, which are more than four points at a set distance threshold apart. Then, using satellite image, street map, grid structure, geographical location of power plants and substations, specific route of transmission lines, land use information and historical load data, commercial professional software of geographic information system is used to establish the registration base map of the area to be predicted. In the registered base map, the area to be predicted is divided according to the different land use types set, and a land use information layer is established; The prediction area is divided according to the power supply range of the 10kV feeder, and a type I cell layer is established. Then, the prediction area is divided according to a regular grid of equal-sized squares with a side length of 300m, and a type II cell layer is established.

3. The prediction method considering the spatiotemporal dual-dimensional characteristics of spatial power load according to claim 2, characterized in that: The different land use types mentioned include residential land, commercial land, cultural and entertainment land, industrial land, administrative office land, green land, municipal facilities land, and special land.

4. The prediction method considering the spatiotemporal dual-dimensional characteristics of spatial power load according to claim 1, characterized in that: The specific method for generating the spatial load spatiotemporal information matrix based on type II cells in step two is as follows: ① The spatial power load gridding technology is used to transform the spatial power load data based on type I cells into spatial power load data based on type II cells at the corresponding time. ② Treat each type II cell as a node, use the load data of each type II cell as the feature information of the node, and use the adjacency relationship between each type II cell as the edge to generate a spatial load spatiotemporal information matrix based on type II cells. This matrix contains the load values ​​of all type II cells and their corresponding spatial location distribution information.

5. The prediction method considering the spatiotemporal dual-dimensional characteristics of spatial power load according to claim 1, characterized in that: The specific method for extracting the spatial dimension features of type II cell loads is as follows: ① Generate adjacency matrix A using the spatial load spatiotemporal information matrix based on type II cells; Two type II cells are considered as two cell nodes. If they share a common boundary in their spatial distribution, then these two type II cells are considered to be adjacent. This is demonstrated using the adjacency matrix A∈R. N×N The connection relationship between type II cells is represented by R, where R is the set of real numbers, N is the total number of type II cells, and two cell nodes are represented by cell node v. p With cell node v q If we express that the value a in the p-th row and q-th column of the adjacency matrix A is... pq =1, Conversely, cell node v p With cell node v q If they are not adjacent, then a pq =0; ② Calculate the attention coefficient; Calculate the target type II cell node v using the attention mechanism. i With neighboring type II cell node v j The attention coefficient is calculated using the formula shown in formula (1): e ij =σ1(α T [Wh i ||Wh j ]) (1); In the formula, e ij For type II cell node v i With neighboring type II cell node v j Attention coefficient; σ1(·) is the Leaky ReLU activation function; α T W is the learnable parameter vector; W is the weight parameter matrix; || denotes vector concatenation; h i h j They are Class II cell nodes v i and v j The characteristics; i and j represent the i-th cell and the j-th cell in all type II cells, respectively; ③ Normalize attention weights; To make the attention coefficients of different nodes comparable, the softmax function is used to normalize the attention coefficients of different nodes, resulting in the normalized attention coefficient α. ij ; ④ Weighted aggregation of neighbor node features; To better learn the feature information between each type II cell, a multi-head attention mechanism is applied to perform a weighted summation of the features of neighboring type II cell nodes. The features of each type II cell node are calculated by K independently working attention mechanisms, and the results are concatenated to obtain the final features of the type II cell nodes. The calculation formula is shown in formula (2). In the formula, h i ′ is a type II cell node v i The features after weighted aggregation; k = 1, 2, ..., K, where K is the number of attention heads; σ²(·) is the sigmoid activation function; j ∈ L i， L i For type II cell node v i The set of all neighboring Class II cells; W represents the normalized attention coefficient for the k-th attention head. k h is the weight parameter matrix for the k-th attention head. j For type II cell node v j Its characteristics.

6. The prediction method considering the spatiotemporal dual-dimensional characteristics of spatial power load according to claim 1, characterized in that: The specific method for extracting the time dimension features of type II cell load in step four is as follows: ①Calculation of dilated causal convolution; The DeepTCN expands the receptive field through the dilation coefficient d to capture the long-term temporal development pattern of type II cell loads; the type II cell load data is temporally unfolded and dilated causal convolution is calculated, as shown in formula (3): In the formula, y t (l) W represents the type II cell feature vector output by the l-th layer of the network at time t; m = 0, 1, ..., M-1, where M is the kernel size; m (l) The m-th weight parameter matrix of the convolutional kernel in the l-th layer of the network; y t-dm (l-1) denoted as , where is the type II cell feature vector output by the (l-1)th layer of the network at time t-dm; d is the inflation coefficient. ② Calculate the time dimension characteristics of type II cell loading; The DeepTCN temporal convolutional network extracts the temporal features of type II cellular loads by combining dilated causal convolution and skip connections. A residual block is formed by two dilated causal convolutions. Stacking multiple residual blocks captures the long-term dependencies of type II cellular loads. Then, skip connections linearly superimpose the input data of type II cellular loads with the output features of the residual blocks to obtain the temporal features of type II cellular loads. The calculation formula for skip connections is shown in Equation (4). and' t / and t +x t (4); In the formula, y′ t y is the time-series feature vector of type II cell load output by the model at time t; t x is the output eigenvector of the residual block at time t; t Let be the original eigenvector of the type II cell load at time t.

7. The prediction method considering the spatiotemporal dual-dimensional characteristics of spatial power load according to claim 1, characterized in that: The specific method for fusing the spatiotemporal features of type II cell loads in step five is as follows: A spatiotemporal fusion layer is constructed using a multi-head attention mechanism to deeply extract the spatial and temporal features of type II cellular loads; First, linear transformations are performed on the spatial and temporal feature matrices of the type II cell load to generate the Q, K, and V matrices required for each attention head. Secondly, for each attention head, the attention weights are obtained by calculating the dot product of matrix Q and matrix K, and then performing scaling and softmax operations. Then, the matrix V is weighted and summed using attention weights to obtain the output features of each head; Finally, the outputs of all attention heads are concatenated and linearly transformed to obtain the fused spatiotemporal feature matrix of type II cell load.

8. The prediction method considering the spatiotemporal dual-dimensional characteristics of spatial power load according to any one of claims 1 to 7, characterized in that: In step six, a prediction model considering the spatiotemporal dual-dimensional characteristics of spatial power load is constructed. The specific method for predicting the spatiotemporal dual-dimensional characteristics of power load is as follows: A prediction model considering the spatiotemporal dual-dimensional features of spatial power load is established using graph attention network (GAT), deep temporal convolutional network (DeepTCN), and multi-head attention mechanism. ① The spatiotemporal matrix of spatial load based on type II cells is used as the input of the prediction model. GAT is used to weight and aggregate the node features of each type II cell with the node features of adjacent cells to extract the spatial dimension features of type II cell load. DeepTCN dilated causal convolution is used to capture long-term dependencies layer by layer. The temporal information of type II cell load is extracted by stacking residual blocks and skip connections to extract the temporal dimension features of type II cell load. ② The spatial and temporal features of the type II cell load are weighted and aggregated using the multi-head attention mechanism of the spatiotemporal fusion layer to obtain the spatiotemporal feature matrix of the type II cell load; ③ The spatial load prediction results based on type II cells are output through the fully connected layer set by the prediction model.

Images