Space load prediction method based on graph convolutional neural network and gating circulation unit

Abstract

The invention discloses a spatial load prediction method based on a graph convolutional neural network and a gating cycle unit, and belongs to the technical field of power distribution network planning. Comprising the steps of integrating basic data information such as a grid structure and land use information of a to-be-predicted area; establishing a registration base map, a land information map layer, a class I cellular map layer and a class II cellular map layer; detecting and correcting singular values in the historical load data of the I-class cells by adopting a GESD (Generally Electro-Static Discharge) algorithm; converting the I-class cellular load into a II-class cellular load by using a space power load gridding technology; and constructing a spatial load prediction model by using the GCN and the GRU, extracting spatial features among the II-type cellular loads by using the GCN, obtaining time sequence features of the II-type cellular loads by using the GRU, and outputting a spatial load prediction result based on the II-type cells through a full connection layer. According to the method, the space influence relation between the cellular loads and the time sequence development rule of the cellular loads are comprehensively considered, the prediction precision is improved, and more accurate data support can be provided for power system planning.

Description

Technical Field

[0001] The present invention belongs to the field of distribution network planning in power systems, and in particular relates to a spatial load prediction method based on a graph convolutional neural network and a gated cyclic unit. Background Art

[0002] Spatial load forecasting is the prediction of the future load size and location distribution in the power supply area, also known as spatial load forecasting (SLF). SLF is the basis of urban power grid planning, and its results will directly affect the rationality and scientificity of power equipment configuration. The research on SLF is of great significance to improving the economy, efficiency and reliability of power system construction. Determining the future load size and spatial distribution of the planned area can more accurately plan and design the substation layout and line channels of the power grid, thereby establishing a reasonable and reliable grid structure.

[0003] Most of the current spatial load forecasting methods focus on mining the temporal characteristics of cellular load data, without fully considering the spatial influence relationship between cells, and the existing spatial load forecasting methods directly use Class I cellular load data for forecasting. Because Class I cells are generated according to the power supply range of power facilities, power supply units, etc., the shape, number, size and location of Class I cells will change with the commissioning of new power equipment in the forecast area, thereby affecting the accuracy of spatial load forecasting.

[0004] In this regard, the present invention proposes a spatial load prediction method based on graph convolution neural networks (GCN) and gated recurrent units (GRU), which uses spatial power loads characterized by class II cells for prediction. Because class II cells are generated according to regular grid divisions of equal size, the generation process is independent of the switching on and off of power equipment and changes in its power supply range, so the shape, quantity, size and position of class II cells will naturally not change, that is, it will not affect the accuracy of spatial load prediction. The spatial load prediction model is constructed using GCN and GRU, which comprehensively considers the spatial influence relationship between cellular loads and the temporal development law of the cellular loads themselves, improves the prediction accuracy, and provides accurate data support for power grid planning. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide a spatial load prediction method based on a graph convolutional neural network and a gated recurrent unit, which is used to solve the disadvantages of insufficient accuracy caused by directly using Class I cellular load data for prediction in the prior art and the problem that the spatial influence between cellular loads is not sufficiently considered in current research.

[0006] The technical solution adopted by the present invention is to provide a spatial load prediction method based on a graph convolutional neural network and a gated recurrent unit, comprising the following steps:

[0007] 1) Establish a power geographic information system based on MapInfo, and establish a registration base map, land use information layer, type I cell layer and type II cell layer;

[0008] 2) The GESD algorithm is used to detect and correct the singular values ​​in the historical load data of type I cells;

[0009] 3) Use spatial power load gridding technology to transform type I cellular load into type II cellular load, and generate a spatial load spatiotemporal information map based on type II cells;

[0010] 4) Use GCN to extract the spatial features between class II cell loads and establish a GCN-based spatial feature extraction model;

[0011] 5) Use GRU to obtain the temporal characteristics of the type II cell load and establish a GRU-based temporal feature extraction model;

[0012] 6) Use GCN and GRU to build a spatial load prediction model, and output the spatial load prediction results based on type II cells through the fully connected layer.

[0013] The specific content of step 1) is:

[0014] ① Collect and integrate the grid structure, land use information, 10kV feeder power supply range and corresponding historical load data within the planning area;

[0015] ② In the electric power geographic information system environment, a registered base map of the area to be planned is constructed; in the registered base map, the area to be planned is divided into eight land use types according to the urban planning scheme, and a land use information layer is constructed. Type I cells are generated according to the power supply range of the 10kV feeder, and a type I cell layer is constructed. Type II cells are generated with equal-sized grids with a given side length, and a type II cell layer is constructed.

[0016] The specific content of step 2) is:

[0017] ① Define the upper limit of the number of outliers in the historical load data set of type I cells as r; the number of cycles is recorded as i, and the initial i=1;

[0018] ② Calculate the average value of the type I cell load data set X and standard deviation δ, where δ is calculated as shown in formula (1),

[0019]

[0020] Where P is the amount of data contained in the type I cell load data set; X j is the data corresponding to the jth type I cell load in the type I cell load data set;

[0021] ③ Calculate the data deviation of each type I cell load separately, and select the largest deviation as the test statistic, denoted as R i , and its calculation formula is shown in formula (2).

[0022]

[0023] Where, X i is the type I cell load value corresponding to the i-th test statistic;

[0024] ④Calculate the test statistic R i The corresponding test critical value is denoted as λ i , see formula (3) and formula (4),

[0025]

[0026] In the formula, t P-i-1,β is the t distribution value with Pi-1 degrees of freedom and β probability level; α is the significance level;

[0027] ⑤The test statistic R i and the test critical value λ i In comparison, if R i >λ i , then the data is an outlier, and proceed to step 6; if R i ≤λ i , then jump to step 7;

[0028] ⑥ Delete the test statistic R in the data set of type I cell loads i The corresponding type I cell load data X i , get a new type I cell load data set; if i = r, all outliers have been detected, and execute step 7; otherwise, i = i + 1 and return to step 2 to continue the cycle;

[0029] ⑦ Output all abnormal values ​​of the type I cell load data set and end the loop.

[0030] The specific content of step 3) is:

[0031] ① Process the outliers of the type I cellular load data detected by the GESD algorithm. For the detected single outlier, use the average value of the load data of the day to replace it. For the detected continuous outlier sequence, use the load data corresponding to similar days to replace it, and obtain the corrected type I cellular load data;

[0032] ② In the power GIS environment, the modified type I cell load data is transformed into spatial load data described by type II cells using spatial load gridding technology;

[0033] ③ Taking the geometric center of each Class II cell as the node, the load data of each Class II cell as the characteristic information of the node, and the structure between each Class II cell as the edge, a spatial load spatiotemporal information graph described by Class II cells is constructed to describe the spatial position relationship between all Class II cells and the load value corresponding to the Class II cells.

[0034] The specific content of step 4) is:

[0035] ① Generate adjacency matrix A and degree matrix D; Generate adjacency matrix A and degree matrix D using the spatial load spatiotemporal information graph described by type II cells, and model the spatial correlation of type II cells; where A∈R N×N (N is the number of type II cells), which is used to represent the connection relationship between type II cells; since the distance between any two adjacent type II cell nodes is equal (in the spatial distribution of type II cells, if two type II cells have a common edge, they are considered to be adjacent), in the spatial load spatiotemporal information diagram of type II cells, if the cell node v i With cell node v j Adjacent, then a ij =1(a ij is the value of the i-th row and j-th column in A), otherwise, if the cell node v i With cell node v j If they are not adjacent, then a ij =0; D∈R N×N is a diagonal matrix, where d ii Represents the type II cell node v i The number of adjacent type II cells is calculated by formula (5).

[0036]

[0037] Where, d ii is the value of the i-th row and i-th column in the degree matrix (i=1,2,…,N);

[0038] ② Normalize the adjacency matrix; add the unit matrix to the adjacency matrix A and the degree matrix D to obtain a new adjacency matrix and the new degree matrix Then reuse right After normalization, we get The calculation formula is shown in formula (6)-(8):

[0039]

[0040]

[0041]

[0042] Where I is the unit matrix;

[0043] ③ Output the spatial characteristic matrix of type II cellular load; obtain the node characteristic matrix of type II cellular load through the spatial load spatiotemporal information graph depicted by type II cells, and the normalized adjacency matrix The node feature matrix of the type II cell is aggregated with the node features of the type II cell load itself and the node features of the adjacent type II cell loads through the GCN hidden layer. The calculation formula is shown in formula (9):

[0044]

[0045] In the formula, H (l) H is the type II cell load node feature matrix input to the lth hidden layer; (l+1) is the type II cell load node feature matrix output by the lth hidden layer; W (l) is the weight matrix of the lth hidden layer; σ1(·) is the activation function.

[0046] The specific content of step 5) is:

[0047] ① Calculate the candidate state vector of the type II cell load at time t The reset gate of GRU is responsible for controlling the weight of the neuron output data feature that is saved in the previous time step, and converting the class II cell load data x at time t into t and the output vector h of GRU at time t-1 t-1 As input, the calculation formula is shown in formula (10) and formula (11),

[0048] r t =σ2(W r [h t-1 ,x t ]+b r ) (10)

[0049]

[0050] In the formula, r t is the state vector of the reset gate at time t; W r is the weight parameter matrix of the reset gate; is the weight parameter matrix of the candidate set; b r , b his the bias vector; [] indicates vector concatenation; σ2(·) is the activation function; tanh(·) is the hyperbolic tangent activation function; W is the Hadamard product operation, that is, the elements in matrices of the same order are multiplied correspondingly;

[0051] ② Calculate the time series characteristics of the class II cell load; the update gate of GRU is responsible for controlling the information weight of the neuron input data at the current time step. The relevant calculation formulas are shown in formula (12) and formula (13).

[0052] z t =σ2(W z [h t-1 ,x t ]+b z ) (12)

[0053]

[0054] In the formula, z t Update the state vector of the gate at time t; W z is the weight parameter matrix of the update gate; b z is the bias vector; h t is the time series characteristic vector of the type II cell load at time t.

[0055] The specific content of step 6) is: the spatial feature extraction model and the temporal feature extraction model together constitute a spatial load prediction model based on type II cells;

[0056] ① The spatial load spatiotemporal information graph data of type II cells and the corresponding adjacency matrix are used as the input of the spatial feature extraction model, and the node features of each type II cell and the node features of adjacent cells are aggregated using the hidden layer to extract the spatial features of the type II cell load;

[0057] ② Input the type II cellular load after extracting spatial features into the time series feature extraction model to extract time series features, and obtain the spatiotemporal feature matrix of the type II cellular load;

[0058] ③ Output the spatial load prediction results based on type II cells through the fully connected layer.

[0059] Through the above design scheme, the present invention can bring the following beneficial effects:

[0060] a) The present invention uses spatial power loads characterized by Class II cells for prediction, which can effectively overcome the problem that the shape, number and size of Class I cells change due to changes in the power supply range or the introduction of new power equipment in the predicted area within a given time period, thereby affecting the SLF.

[0061] b) The present invention uses the GESD algorithm to detect and process singular values ​​of Class I cell load historical data, which can effectively eliminate the adverse effects of singular values ​​in the load data on the SLF.

[0062] c) The present invention comprehensively considers "the spatial influence relationship between the type II cells mined by GCN" and "the temporal development law of the type II cell load mined by GRU", and has higher prediction accuracy than other typical SLF methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 It is a basic principle diagram of the spatial load prediction method based on graph convolutional neural network and gated recurrent unit of the present invention;

[0064] Figure 2 It is a land use information graph and a type I cell graph of the area to be planned of the spatial load prediction method based on graph convolutional neural network and gated cyclic unit of the present invention;

[0065] Figure 3 The invention provides a spatial load prediction method based on graph convolutional neural network and gated recurrent unit, and a power load spatial feature extraction model based on GCN and characterized by type II cells;

[0066] Figure 4 The invention provides a spatial load forecasting method based on graph convolutional neural network and gated recurrent unit, and a power load time series feature extraction model characterized by type II cells based on GRU;

[0067] Figure 5 A diagram of a spatial load prediction model based on type II cells of a spatial load prediction method based on a graph convolutional neural network and a gated recurrent unit according to the present invention;

[0068] Figure 6 It is the quasi-measured value and predicted value of the class II cell load of the spatial load prediction method based on graph convolutional neural network and gated recurrent unit of the present invention;

[0069] Figure 7 It is a comparison diagram of the amplitude errors of the type II cellular load prediction results of the spatial load prediction method based on graph convolutional neural network and gated recurrent unit of the present invention;

[0070] Figure 8 This is an error comparison diagram of the class II cellular load prediction results of the spatial load prediction method based on graph convolutional neural network and gated recurrent unit of the present invention. DETAILED DESCRIPTION

[0071] The present invention is further described below with reference to the accompanying drawings and specific embodiments:

[0072] The present invention provides a spatial load prediction method based on graph convolutional neural network and gated recurrent unit, the basic principle diagram is as follows Figure 1 As shown, the example contents of the method include: establishing a power geographic information system (GIS), data preprocessing, building a spatial load forecasting model, and error evaluation and analysis.

[0073] In the scenario where the "Five-Year Plan for Power Grid Development Planning" is revised on a rolling basis every year, according to the spatial load prediction method based on graph convolutional neural network and gated cyclic unit proposed in the present invention, taking some administrative regions of a city in Northeast my country as an example, using its historical load data values ​​from 2009 to 2013, the spatial load value of the region in 2014 (the annual maximum value of the load of each Class II cell) is predicted. In order to verify the accuracy of the method proposed in the present invention, the prediction results of the method of the present invention are compared and analyzed with the prediction results of the traditional grey theory method, linear regression method, exponential smoothing method, GCN prediction method, GRU prediction method, GCN-LSTM prediction method, and STGAT prediction method. The specific contents are as follows:

[0074] 1) Establishing power GIS

[0075] Under the base map in the power GIS, according to the urban planning scheme, the land use types of the area to be predicted are divided into eight categories: residential land, commercial land, cultural and entertainment land, industrial land, administrative office land, green land, municipal facilities land and special land. A land use information layer is established, and a type I cell layer is generated according to the 10kV power supply range, and a type II cell layer is generated according to a regular square grid of equal size with a side length of 300m. Figure 2 It is a land use information map and a type I cell map.

[0076] 2) Data preprocessing

[0077] The sampling interval of the measured values ​​of the historical load data of the Class I cells corresponding to each 10kV feeder in the prediction area is 5 minutes. The historical load data of the Class I cells are grouped by month, and the GESD (generalized extreme studentized deviate) algorithm is used to detect the singular values ​​in each group of historical load data; the singular values ​​detected by the GESD algorithm are corrected by replacing the single singular value with the average load value of the day and replacing the continuous abnormal values ​​with the load data of the corresponding period of similar days.

[0078] Since the spatial load forecasting method proposed in the present invention mainly serves the power system planning, the maximum load value of each month in each Class I cell load from 2009 to 2013 is extracted as the reasonable maximum monthly load of the Class I cell, and in the power GIS, the spatial load gridding technology is used to convert it into the reasonable maximum value of the Class II cell load, and then the geometric center of each Class II cell is used as the node, the Class II cell load data is used as the characteristic information of the node, and the structure between Class II cells is used as the edge to construct the spatial load spatiotemporal information graph depicted by Class II cells as the input of the forecasting model. By predicting the load values ​​of each month of Class II cells in 2014, the annual maximum load of Class II cells is determined and used as the output of the forecasting model.

[0079] 3) Constructing a spatial load prediction model

[0080] Use GCN to build a spatial feature extraction model. The model structure diagram is as follows Figure 3 As shown in the figure, the model contains multiple hidden layers, each of which only processes the first-order neighborhood feature information. The superposition of multiple hidden layers can realize the transmission and aggregation of multi-order neighborhood feature information. GRU is used to build a temporal feature extraction model. The model structure diagram is shown in the figure Figure 4 As shown in the figure, GRU has two gate structures: reset gate and update gate. The reset gate is responsible for controlling the weight of the neuron output data feature of the previous time step, and the update gate is responsible for controlling the weight of the neuron input data information of the current time step. These weights are updated through the training set during the training process. After the training is completed, the weights are saved to form a time series feature extraction model.

[0081] The prediction model of the SLF method of the present invention is composed of a spatial feature extraction model and a temporal feature extraction model. The model structure diagram is shown in FIG. Figure 5 As shown in the figure. The parameters of the prediction model are as follows: the number of network layers of GCN is 2; the number of hidden layer units is 32; the number of training iterations is 100; the learning rate is 0.001; the number of network layers of GRU is 1; the number of hidden layer units is 64; the number of training iterations is 100; the batch size is 32; the input layer time step is 5; the learning rate is 0.001; the model uses the mean square error loss function; GCN uses Dropout set to 0.2 to prevent the model from overfitting; GRU uses L2 regularization to prevent the model from overfitting, and the weight decay coefficient is 0.01. The spatial load spatiotemporal information graph of the type II cell is first obtained through the spatial feature extraction model to obtain the spatial relationship between the type II cell loads, and the spatial feature matrix of the type II cell is output. Then, the temporal feature law of the type II cell itself is obtained by the temporal feature extraction model, and finally the spatial load prediction result of the type II cell is output through the fully connected layer to obtain the predicted value of the type II cell load in 2014.

[0082] 4) Error evaluation analysis

[0083] In order to more intuitively display the prediction results of type II cells, the quasi-measured values ​​of type II cells and the predicted values ​​of type II cells of five methods other than the traditional method (grey theory method, linear regression method and exponential smoothing method) are displayed in the power GIS, as shown in Figure 2. Figure 6 In the power GIS, the size of all type II cellular load data can be represented by the depth of color. The darker the color, the larger the cellular load data, and vice versa.

[0084] In order to more fully analyze the prediction effect of different spatial load forecasting methods, the amplitude of the prediction result error and the spatial distribution are considered comprehensively, and the amplitude error: average relative error and average absolute error, and the amplitude-space characteristic error: error influence value are used as evaluation indicators for analysis. The prediction errors of type II cells of each method are shown in Table 1, Table 2, Figure 7 and Figure 8 shown.

[0085] Table 1 Amplitude error of load prediction value of type II cell

[0086]

[0087] It can be concluded from Table 1 that, compared with the GCN prediction method and the GRU prediction method, the average absolute error of the method of the present invention is reduced by 0.018MW and 0.021MW, respectively, and the average relative error is reduced by 2.42% and 4%, respectively. This shows that whether it is a prediction method that only considers the spatial influence relationship between cellular loads or only considers the temporal development law of the cellular load itself, it is inferior to the method of the present invention, that is, comprehensively considering the influence of both spatial and temporal aspects of cellular loads can effectively improve the prediction accuracy.

[0088] As shown in Table 1, the average absolute error and average relative error of the GCN prediction method and the GRU prediction method are higher than those of the GCN-LSTM prediction method and the STGAT prediction method, which further shows that it is not sufficient to only consider the spatial relationship between cellular loads or the temporal development law of cellular loads when performing spatial load forecasting. Compared with the GCN-LSTM prediction method, the average absolute error and average relative error of the method of the present invention are 0.004MW and 0.36% lower, respectively. It can be seen that the method of the present invention is better than the GCN-LSTM prediction method, which also shows that GRU is better than LSTM when processing time series.

[0089] Figure 7 The figure is a comparison of the amplitude errors of the prediction results of eight SLF methods, where the shaded parts in sub-figures (a) and (b) are circles with the average absolute error and average relative error of the prediction results obtained by the method of the present invention as the radius. Obviously, the SLF prediction accuracy of the prediction results whose amplitude error values ​​fall outside the shaded circle is lower than that of the method of the present invention.

[0090] Depend on Figure 7 It can be seen that compared with the artificial intelligence prediction methods (for example, the method of the present invention, the GCN prediction method, the GRU prediction method, the GCN-LSTM prediction method, and the STGAT prediction method), the average absolute error and the average relative error of the prediction results of the traditional prediction methods (for example, the gray theory method, the linear regression method, and the exponential smoothing method) are farther away from the origin, that is, the amplitude error of the traditional methods is generally larger, indicating that SLF using artificial intelligence technology can significantly improve the prediction accuracy.

[0091] Depend on Figure 7 (a) It can be seen that the average absolute errors of the prediction results obtained by the seven prediction methods except the method of the present invention are all outside the shaded circle. Obviously, the prediction accuracy of the method of the present invention is the highest. Figure 7 (b) It can be seen that only the average relative errors of the method of the present invention and the STGAT prediction method are within the shaded circle, and both are better than the other six SLF methods. However, the average relative error of the STGAT prediction method is closer to the origin. It can be seen that it is difficult to judge the advantages and disadvantages of the two methods only by comparing the amplitude errors.

[0092] Table 2 Amplitude-space characteristic error of type II cell load prediction value

[0093]

[0094] Table 2 shows the amplitude-space errors of all prediction methods. It can be concluded from Table 2 that, compared with other comparison methods, the error impact value index of the method of the present invention is the smallest, that is, the adverse impact on power grid planning is the smallest. Figure 8 By comparing the three error evaluation indicators of the prediction results of the method of the present invention with those of the other seven methods, it can be seen that the prediction effects of the STGAT prediction method and the method of the present invention are better than those of the other six prediction methods. Overall, the prediction effect of the method of the present invention is better than that of the STGAT prediction method, that is, the prediction accuracy of the method of the present invention is the highest, and it can provide more accurate data support for power grid planning.

[0095] The implementation methods of the present invention are not limited to the above-mentioned embodiments, and any other changes, modifications, substitutions, combinations, and simplifications made without departing from the spirit and principle of the present invention should be equivalent replacement methods and are included in the protection scope of the present invention.

Claims

1. A spatial load prediction method based on graph convolutional neural network and gated recurrent unit, characterized by: The following steps are involved: 1) Establish a power geographic information system based on MapInfo, and establish a registration base map, land use information layer, type I cell layer and type II cell layer; 2) The GESD algorithm is used to detect and correct the singular values ​​in the historical load data of type I cells; 3) Use spatial power load gridding technology to transform type I cellular load into type II cellular load, and generate a spatial load spatiotemporal information map based on type II cells; 4) Use GCN to extract the spatial features between class II cell loads and establish a GCN-based spatial feature extraction model; 5) Use GRU to obtain the temporal characteristics of the type II cell load and establish a GRU-based temporal feature extraction model; 6) Use GCN and GRU to build a spatial load prediction model, and output the spatial load prediction results based on type II cells through the fully connected layer.

2. The spatial load prediction method based on graph convolutional neural network and gated recurrent unit according to claim 1, characterized in that: The specific content of step 1) is: ① Collect and integrate the grid structure, land use information, 10kV feeder power supply range and corresponding historical load data within the planning area; ② In the electric power geographic information system environment, a registered base map of the area to be planned is constructed; in the registered base map, the area to be planned is divided into eight land use types according to the urban planning scheme, and a land use information layer is constructed. Type I cells are generated according to the power supply range of the 10kV feeder, and a type I cell layer is constructed. Type II cells are generated with equal-sized grids with a given side length, and a type II cell layer is constructed.

3. The spatial load prediction method based on graph convolutional neural network and gated recurrent unit according to claim 1, characterized in that: The specific content of step 2) is: ① Define the upper limit of the number of outliers in the historical load data set of type I cells as r; the number of cycles is recorded as i, and the initial i=1; ② Calculate the average value of the type I cell load data set and standard deviation δ, where δ is calculated as shown in formula (1), Where P is the amount of data contained in the type I cell load data set; X j is the data corresponding to the jth type I cell load in the type I cell load data set; ③ Calculate the data deviation of each type I cell load separately, and select the largest deviation as the test statistic, denoted as R i , and its calculation formula is shown in formula (2). Where, X i is the type I cell load value corresponding to the i-th test statistic; ④Calculate the test statistic R i The corresponding test critical value is denoted as λ i , see formula (3) and formula (4), Where, t P-i-1,β is the t distribution value with Pi-1 degrees of freedom and β probability level; α is the significance level; ⑤The test statistic R i and the test critical value λ i In comparison, if R i >λ i , then the data is an outlier, and proceed to step 6; if R i ≤λ i , then jump to step 7; ⑥ Delete the test statistic R in the data set of type I cell loads i The corresponding type I cell load data X i , get a new type I cell load data set; if i = r, then all outliers have been detected, and execute step 7; Otherwise, i=i+1 and return to step 2 to continue the loop; ⑦ Output all abnormal values ​​of the type I cell load data set and end the loop.

4. The spatial load prediction method based on graph convolutional neural network and gated recurrent unit according to claim 1, characterized in that: The specific content of step 3) is: ① Process the outliers of the type I cellular load data detected by the GESD algorithm. For the detected single outlier, use the average value of the load data of the day to replace it. For the detected continuous outlier sequence, use the load data corresponding to similar days to replace it, and obtain the corrected type I cellular load data; ② In the power GIS environment, the modified type I cell load data is transformed into spatial load data described by type II cells using spatial load gridding technology; ③ Taking the geometric center of each Class II cell as the node, the load data of each Class II cell as the characteristic information of the node, and the structure between each Class II cell as the edge, a spatial load spatiotemporal information graph described by Class II cells is constructed to describe the spatial position relationship between all Class II cells and the load value corresponding to the Class II cells.

5. The spatial load prediction method based on graph convolutional neural network and gated recurrent unit according to claim 1, characterized in that: The specific content of step 4) is: ① Generate adjacency matrix A and degree matrix D; Generate adjacency matrix A and degree matrix D using the spatial load spatiotemporal information graph described by type II cells, and model the spatial correlation of type II cells; where A∈R N×N (N is the number of type II cells), which is used to represent the connection relationship between type II cells; since the distance between any two adjacent type II cell nodes is equal (in the spatial distribution of type II cells, if two type II cells have a common edge, they are considered to be adjacent), in the spatial load spatiotemporal information diagram of type II cells, if the cell node v i With cell node v j Adjacent, then a ij =1(a ij is the value of the i-th row and j-th column in A), otherwise, if the cell node v i With cell node v j If they are not adjacent, then a ij =0; D∈R N×N is a diagonal matrix, where d ii Represents the type II cell node v i The number of adjacent type II cells is calculated by formula (5). Where, d ii is the value of the i-th row and i-th column in the degree matrix (i=1,2,…,N); ② Normalize the adjacency matrix; Add the identity matrix to the adjacency matrix A and the degree matrix D to get the new adjacency matrix and the new degree matrix Then reuse right After normalization, we get The calculation formula is shown in formula (6)-(8): Where I is the unit matrix; ③ Output the spatial characteristic matrix of type II cell load; The node characteristic matrix of type II cellular load is obtained by the spatial load spatiotemporal information graph depicted by type II cells. The normalized adjacency matrix The node feature matrix of the type II cell is aggregated with the node features of the type II cell load itself and the node features of the adjacent type II cell loads through the GCN hidden layer. The calculation formula is shown in formula (9): In the formula, H (l) is the type II cell load node feature matrix input to the lth hidden layer; H (l+1) is the type II cell load node feature matrix output by the lth hidden layer; W (l) is the weight matrix of the lth hidden layer; σ1(·) is the activation function.

6. The spatial load prediction method based on graph convolutional neural network and gated recurrent unit according to claim 1, characterized in that: The specific content of step 5) is: ① Calculate the candidate state vector of the type II cell load at time t The reset gate of GRU is responsible for controlling the weight of the neuron output data feature that is saved in the previous time step, and converting the class II cell load data x at time t into t and the output vector h of GRU at time t-1 t-1 As input, the calculation formula is shown in formula (10) and formula (11), r t =σ2(W r [h t-1 ,x t ]+b r ) (10) In the formula, r t is the state vector of the reset gate at time t; W r is the weight parameter matrix of the reset gate; is the weight parameter matrix of the candidate set; b r 、b h is the bias vector; [] indicates vector concatenation; σ2(·) is the activation function; tanh(·) is the hyperbolic tangent activation function; W is the Hadamard product operation, that is, the elements in matrices of the same order are multiplied correspondingly; ② Calculate the time series characteristics of type II cell load; The update gate of GRU is responsible for controlling the weight of the neuron input data information at the current time step. The relevant calculation formulas are shown in formula (12) and formula (13). z t =σ2(W z [h t-1 ,x t ]+b z ) (12) In the formula, z t Update the state vector of the gate at time t; W z is the weight parameter matrix of the update gate; b z is the bias vector; h t is the time series characteristic vector of the type II cell load at time t.

7. The spatial load prediction method based on graph convolutional neural network and gated recurrent unit according to claim 1, characterized in that: The specific content of step 6) is: the spatial feature extraction model and the temporal feature extraction model together constitute a spatial load prediction model based on type II cells; ① The spatial load spatiotemporal information graph data of type II cells and the corresponding adjacency matrix are used as the input of the spatial feature extraction model, and the node features of each type II cell and the node features of adjacent cells are aggregated using the hidden layer to extract the spatial features of the type II cell load; ② Input the type II cellular load after extracting spatial features into the time series feature extraction model to extract time series features, and obtain the spatiotemporal feature matrix of the type II cellular load; ③ Output the spatial load prediction results based on type II cells through the fully connected layer.

Images