A regional electric power prediction method, system, device and storage medium

Through a three-dimensional convolutional neural network combining attention mechanism, the spatial and temporal characteristics of power nodes are extracted using electrical distance and proximity to solve the problems of complex nonlinear relationship capture and model adjustment difficulty in the existing technology, and the target area is achieved Efficient prediction of power.

CN118657408BActive Publication Date: 2024-11-26SHANDONG UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411139259.4
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-08-20
Publication Date
2024-11-26
Estimated Expiration
2044-08-20

AI Technical Summary

Technical Problem

The prior art cannot effectively capture complex nonlinear relationships and the increase in the number of hyperparameters caused by model cascades, and the difficulty of model adjustment and optimization is increased, limiting the universality and efficiency of regional electrical power prediction.

Method used

A three-dimensional convolutional neural network that combines attention mechanism is used to spatially mesh the power nodes in the target area, calculate the electrical distance and proximity center, assign attention weights, and input the three-dimensional convolutional network to extract space-time dynamic features. , predict future power of the target area.

Benefits of technology

Effectively capture complex nonlinear relationships under low model complexity, reducing the difficulty of model adjustment and optimization, and achieving comprehensive prediction of the power of the target area.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118657408B_ABST
    Figure CN118657408B_ABST
Patent Text Reader

Abstract

The invention discloses a method, system, device and storage medium for regional electric power prediction, and relates to the field of energy technology. The method comprises: spatially gridding a target area to obtain a plurality of grids containing only a single power node; synchronizing the time of each grid node; calculating the electrical distance between electrically associated nodes, and dividing the electrically associated nodes into a power path according to the electrical distance; calculating the proximity centrality of each node in each power path based on the electrical distance; allocating attention weights to the power nodes according to the electrical distance and the proximity centrality; arranging the matrix obtained after processing in the order of timestamps, and using a three-dimensional convolutional network with different time granularities to predict the power of the target area at a future moment; the method adopts a three-dimensional convolutional neural network combined with an attention mechanism to perform power prediction, extracts the spatiotemporal characteristics of power data under low model complexity, and further predicts the power of the target area.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of energy technology, and in particular to a regional electric power prediction method, system, device and storage medium. Background Art

[0002] In the era of energy diversification, regional power forecasting has gone beyond the scope of a single energy type. It covers the comprehensive forecast of power generation from traditional and renewable energy to terminal power consumption such as household electricity. This comprehensive forecasting capability is essential to adapt to the challenges currently facing the power system. Today, the large-scale integration of renewable energy has brought volatility and uncertainty to the power grid, while the diversification and personalized needs of terminal energy consumption patterns such as household electricity have also put forward higher requirements for the balance of supply and demand of the power grid. Comprehensive power forecasting can not only help grid operators better manage the stable operation of the power grid and ensure the effective distribution of energy, but also meet the personalized needs of end users for energy management and improve the efficiency and satisfaction of energy use.

[0003] Current regional electric power forecasting research is mainly carried out independently in the fields of photovoltaic power generation, wind power generation and household energy demand. When performing these forecasting tasks, models are usually selected in advance based on data characteristics, and their respective characteristics are taken into account in the process of model selection and adjustment. This method relies on manual pre-feature analysis to guide the forecasting process. However, this method limits the universality of the model in global power forecasting tasks. When other data need to be predicted, the model needs to be retrained or even replaced, and it is impossible to effectively predict various powers in the target area;

[0004] The main methods currently used in the field of power prediction are divided into two categories. One is based on traditional statistical methods for prediction. This method is simple, easy to implement, and has strong interpretability, but it cannot effectively capture complex nonlinear relationships. The other is the combination of multiple deep learning models, such as CNN and LSTM for prediction. This method can capture complex nonlinear relationships, but due to the increase in the number of hyperparameters caused by model cascading, the difficulty of model adjustment and optimization also increases accordingly. Summary of the invention

[0005] In view of the shortcomings of the existing technology that it is unable to effectively capture complex nonlinear relationships and that the number of hyperparameters increases due to model cascading, and the difficulty of model adjustment and optimization increases accordingly, the present invention proposes a regional electric power prediction method, system, device and storage medium. By adopting a three-dimensional convolutional neural network combined with an attention mechanism to perform power prediction, the spatiotemporal characteristics of power data are extracted under low model complexity, and the power of the target area is predicted, thereby solving the problems existing in the existing technology.

[0006] A method for regional electric power prediction comprises the following steps:

[0007] Collect all power nodes in the target area that contain power absorbing or transmitting devices;

[0008] Extract the spatiotemporal dynamic characteristics of each power node;

[0009] Predict the power of the target area at future times based on the spatiotemporal dynamic characteristics of each power node;

[0010] The step of extracting the spatiotemporal dynamic characteristics of each power node specifically includes the following steps:

[0011] The target area is spatially gridded to obtain multiple grids containing only a single power node;

[0012] An equivalent impedance method is used to obtain the electrical distance between the electrically associated power nodes, and the electrically associated power nodes are divided into a power path according to the electrical distance;

[0013] Calculate the closeness centrality of each power node in each power path based on electrical distance;

[0014] Assign attention weights to power nodes based on electrical distance and proximity centrality;

[0015] The power node data after allocating attention weights are input into a three-dimensional convolutional network with different time granularities to obtain the spatiotemporal dynamic characteristics of the power nodes.

[0016] Furthermore, the spatial gridding of the target area to obtain multiple grids containing only a single power node specifically includes the following steps:

[0017] Divide the target area into grids of different sizes according to the power density at different locations; select the largest grid to divide the area with sparse power node distribution and concentrated power density; use the smallest grid as the unit grid, and divide the other grids according to the unit grid size;

[0018] Power nodes are allocated to each grid to obtain a two-dimensional grid containing power nodes; each grid contains only a single power node; if there is a power node spanning multiple grids, the power node is divided into one of the grids that does not contain other power nodes according to the minimum center distance standard.

[0019] Furthermore, the method of using the equivalent impedance method to obtain the electrical distance between the electrically associated power nodes specifically includes the following steps:

[0020] For power nodes in a single power network in a target area, the modulus of the equivalent impedance between the power nodes is used as the electrical distance between the power nodes;

[0021] For the power nodes in the multi-power network in the target area, the multi-power network is split into several single-power networks through the downstream tracing method to obtain the impedance between the power nodes under the action of any power source; after the equivalent impedances between the corresponding nodes under the action of several power sources are connected in parallel, the modulus value is taken as the electrical distance between the power nodes.

[0022] Furthermore, the calculation of the proximity centrality of each power node in each power path based on the electrical distance is expressed as follows:

[0023]

[0024]

[0025] in, Representation Node i The average value of the electrical distance to each point with which there is a power transfer relationship, N Representation and Node i The total number of nodes on the power path, Representation Node i With Node j The electrical distance between Representation Node i The closeness centrality.

[0026] Furthermore, the allocating attention weights to the power nodes according to the electrical distance and the proximity centrality specifically includes the following steps:

[0027] The electrical distance and proximity centrality are normalized, and the electrical distance between two power nodes on a power path after normalization is recorded as ; If there is a node that is electrically associated with multiple nodes, the average of the multiple electrical distances is used;

[0028] The value of the closeness centrality in the power path is denoted by ;

[0029] Assign weights to electrical distance and proximity centrality; if the electrical distance weight is , then the proximity centrality weight is ,and , then the attention weight of the power node is .

[0030] The present invention also includes a regional electric power prediction system, comprising:

[0031] A collection module, used to collect all power nodes in the target area containing devices that absorb power or send power;

[0032] An extraction module, used to extract the spatiotemporal dynamic characteristics of each power node;

[0033] A prediction module is used to predict the power of the target area at a future time according to the spatiotemporal dynamic characteristics of each power node;

[0034] Wherein, the extraction module comprises:

[0035] A division unit is used to spatially grid the target area to obtain multiple grids containing only a single power node;

[0036] An electrical distance calculation unit, used for obtaining the electrical distance between the power nodes with electrical association by using an equivalent impedance method, and dividing the power nodes with electrical association into a power path according to the electrical distance;

[0037] A closeness centrality calculation unit, used to calculate the closeness centrality of each power node in each power path based on the electrical distance;

[0038] an allocation unit, for allocating attention weights to power nodes according to electrical distance and proximity centrality;

[0039] The extraction unit is used to input the power node data after the attention weights are assigned into a three-dimensional convolutional network with different time granularities to obtain the spatiotemporal dynamic characteristics of the power nodes.

[0040] The present invention also includes a regional electric power prediction computer device, including: a memory, a processor and a computer program stored in the memory, and the processor implements the steps of the regional electric power prediction method when executing the computer program.

[0041] Furthermore, a readable storage medium is provided, wherein the readable storage medium stores a computer program, wherein the computer program includes program instructions, and when the program instructions are executed by a processor, they are used to execute the steps of the regional electric power prediction method.

[0042] The present invention provides a regional electric power prediction method, system, device and storage medium, which have the following beneficial effects:

[0043] The present invention divides all power nodes in the target area into grids of different sizes to ensure effective gridding of power nodes in all directions in the target area; uses electrical distance to measure the connection tightness between power nodes, and finds out power nodes with electrical correlation according to the electrical distance, and divides them into a power path. By focusing on analyzing the power changes of nodes on the same power path, it is helpful to reveal the inherent connection and evolution law of power dynamics on a specific power path, and better predict the change of node power. Attention weights are allocated to power nodes according to electrical distance and proximity centrality, and the feature learning ability of three-dimensional convolutional neural network on time scale and space scale is used to extract the spatiotemporal dynamic characteristics of power node data after allocating attention weights under low model complexity, so as to predict the power of the area at future times; the method comprehensively considers various types of power in the target area, effectively captures complex nonlinear relationships, and also reduces the difficulty of model adjustment and optimization, thereby realizing comprehensive prediction of the power of the target area. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 Schematic diagram of effective gridding of target area in an embodiment of the present invention; Figure 1 a represents a distribution diagram of each power node in the area where power prediction is required; Figure 1 b represents a schematic diagram of grid distribution of different sizes divided according to the density of power nodes in different parts of the target area; Figure 1 c represents a schematic diagram of standardized distribution of grids with the smallest grid as the unit grid; Figure 1 The d represents the schematic diagram of allocating power nodes to the grid nearby; Figure 1 e represents the distribution diagram after removing the rows and columns without power nodes; Figure 1 f represents the distribution diagram of the target area after effective gridding;

[0045] Figure 2 This is a schematic diagram of attention weight allocation in an embodiment of the present invention;

[0046] Figure 3 Schematic diagram of a three-dimensional convolutional neural network in an embodiment of the present invention;

[0047] Figure 4 The present invention is a flowchart of a method for predicting regional electric power in an embodiment of the present invention. DETAILED DESCRIPTION

[0048] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0049] The present invention proposes a regional electric power prediction method, which extracts the spatiotemporal characteristics of power data under low model complexity, and then predicts the power of the target area, such as Figure 4 As shown; specifically including the following steps:

[0050] S1. For the target area (which can be any area containing power absorbing or power transmitting equipment, such as a photovoltaic power plant, a town, a community, etc., Figure 1 a) Perform spatial meshing operations and perform mesh divisions of different sizes according to the power node density at different locations in the target area to ensure effective meshing of power nodes in all directions within the target area.

[0051] For areas with sparse power node distribution and low power density, a larger grid can be used; for areas with dense power node distribution and high power density, a smaller grid can be used, such as Figure 1 As shown in b.

[0052] In this data distribution mode, each grid is designed to contain only a single power node, and there is no situation where two or more power nodes appear in one grid; there may be a power node that spans multiple grids, and then it can be divided into one of the grids that does not contain other power nodes through the minimum center distance standard. The gridding that meets the above standards is called effective gridding.

[0053] After the above processing, an initial two-dimensional grid table will be obtained. At this time, the grid table is not neat and regular. Figure 1 As shown in b.

[0054] S2. Take the smallest grid as the unit grid and divide the other grids according to the unit grid size. If the remaining part is less than one unit grid, fill it with one unit grid; Figure 1 As shown in c, Figure 1 b to Figure 1 The grid padding operation is performed during the transformation of c.

[0055] S3. Allocate the power nodes to the nearest grid to form a two-dimensional grid containing the power nodes.

[0056] In the case where the power node is completely within a certain grid, it can be directly assigned to the grid to form a two-dimensional grid containing the power node, and the subsequent sampling data is placed in the position of the grid.

[0057] For the case where the power node spans across the grid, the minimum center distance is used as the basis for determining the grid affiliation. If the grid selected according to the minimum center distance standard contains other power nodes, the grid with the next smallest center distance is selected, and so on. If the power node has the same center distance for several grids, one of the grids can be randomly selected.

[0058] For any power node, its data will be uniquely mapped to a grid, but this does not guarantee that all grids will be assigned data. In other words, there may be some grids that are not filled with data, but all existing power node data will have a corresponding grid. Figure 1 As shown in d, the dark grid area represents the nearest grid to which the power node belongs.

[0059] S4, remove the rows and columns that do not contain power nodes in the two-dimensional data grid to form a new two-dimensional data grid, such as Figure 1 e to Figure 1 The f represents the removal process.

[0060] Given that the interdependence between power nodes in the power system does not significantly depend on their geographical proximity, in order to reduce the data processing complexity and data dimensionality in subsequent analysis, the rows and columns that do not contain power nodes are deleted, reducing the dimension of the original data while retaining some geographical location information.

[0061] S5. Treat each grid area containing a power node as an IoT node and allocate certain communication, computing and storage capabilities to it.

[0062] S6. Use IoT technology to achieve microsecond-level time synchronization and millisecond-level data transmission for all nodes.

[0063] S7. Select a corresponding sampling frequency according to the expected prediction accuracy, fill the collected power value into the corresponding grid, collect power data for a period of time, and add a timestamp to it.

[0064] When recording data, for nodes that send power, the relevant power values ​​are recorded as positive values; for nodes that absorb power, the relevant power data are recorded as negative values.

[0065] The sampling frequency should be at least of the same order of magnitude as the prediction accuracy to ensure the desired prediction accuracy. For this study, due to the limitations of millisecond time transmission technology, power prediction at millisecond time granularity and above can be achieved.

[0066] S8. Use the equivalent impedance method to calculate the electrical distance between nodes that are electrically related, i.e., have power (electrical energy) flow relationships; electrical distance is an indicator used to measure the tightness of the connection between nodes in an electrical network. The smaller the electrical distance between nodes, the tighter the connection between them. Calculate the electrical distance between nodes and use it as one of the bases for the subsequent allocation of attention weights in the convolutional neural network.

[0067] For a single-power network, the modulus of the equivalent impedance between nodes is directly used as the electrical distance between nodes. For a multi-power network, it is split into several single-power networks by downstream tracing (i.e., identifying the current path according to the actual flow path of the current in the electrical system), and the electrical distance is calculated according to the processing method of the single-power network; then, the impedance between each node under the action of any power source can be obtained according to the flow tracing (i.e., calculating and analyzing the voltage, current and power flow of each node in the power system according to the identified current path); for the synthetic flow, the equivalent impedances between the corresponding nodes when several power sources act separately are connected in parallel, and the modulus is taken as the electrical distance between the nodes.

[0068] S9. Divide the nodes with electrical association into one power path.

[0069] According to the electrical distance, we can find out the power nodes with electrical correlation and divide them into one power path. By focusing on analyzing the power changes of nodes on the same power path, it helps to reveal the inherent connection and evolution law of power dynamics on a specific power path and better predict the changes of node power.

[0070] S10. Calculate the closeness centrality of each node in each path.

[0071] Closeness centrality is a metric used to measure the importance of nodes in a network. It is defined as the inverse of the average distance from a node to other nodes. The shorter the distance, the higher the closeness centrality of the node. In a network, nodes with high closeness centrality are usually considered key nodes, with greater control and influence over other nodes. The closeness centrality of each node is calculated using electrical distance and used as one of the bases for the subsequent allocation of attention weights in the convolutional neural network.

[0072] The calculation formula for closeness centrality is as follows:

[0073]

[0074]

[0075] In the above formula, Representation Node i The average value of the electrical distance to each point with which there is a power transfer relationship,N Representation and Node i The total number of nodes on the power path, Representation Node i With Node j The electrical distance between Representation Node i The closeness centrality.

[0076] S11. Allocate attention weights to power nodes based on electrical distance and proximity centrality.

[0077] Different attention maps are generated according to different power paths. The size of the attention map is the same as the input matrix size, and attention weights are assigned to the corresponding positions of the attention map along the power path. The attention weight of each power node on the power path can be determined as follows.

[0078] First, the electrical distance and the closeness centrality index are normalized. The electrical distance between two nodes on a power path after normalization is: For a node that is electrically connected to multiple nodes, the average of multiple electrical distances is used; the closeness centrality value in the power path is Then, weights are assigned to the two indicators of electrical distance and closeness centrality. If the electrical distance weight is , then the proximity centrality weight is ,in The final attention weight of the node is .

[0079] like Figure 2 As shown, an attention weight distribution is given. Figure 2 Lines of different thicknesses represent different power paths; the thickness of the line represents the electrical distance between nodes, and the thicker the line, the closer the electrical distance between nodes.

[0080] S12. Arrange the processed matrix in timestamp order and use a three-dimensional convolutional network with different time granularities to predict the power of the target area at future moments.

[0081] A three-dimensional convolutional neural network is constructed, and the feature learning ability of the three-dimensional convolutional neural network in time and space scales is used to capture the spatiotemporal dynamic characteristics of power data, so as to predict the power at future moments based on the spatiotemporal characteristics of the data.

[0082] like Figure 3 A convolutional network is shown as follows: Assume that the input matrix 1 has a dimension of 4×4×4, which represents a power sampling sequence with three time steps, where each time step contains a 4×4 numerical matrix.

[0083] First, the attention mechanism is introduced. By adding a 4×4 attention layer, the importance or weight of different regions in the input matrix 1 is adjusted to obtain matrix 2. Then, the time and space information in matrix 2 is extracted through the 2×2×2 convolution kernel 1 to obtain matrix 3. A 2×2 pooling layer is used to reduce the dimension of matrix 3 while retaining the most important features to obtain matrix 4. A 1×1×3 convolution kernel 2 is used to further extract the time information to obtain matrix 5. Then, nonlinearity is introduced through the activation layer, and it is mapped to the target dimension through the fully connected layer. Finally, residual calculation and weight update are performed.

[0084] In addition, by constructing three-dimensional convolutional networks with different time granularities, the influence of cumulative prediction errors on the final prediction results can be effectively reduced. The present invention also proposes a feasible solution to construct three-dimensional convolutional neural networks with time granularities of 1s, 5s, and 10s respectively. A network with a time granularity of 10s can reduce the cumulative prediction errors of networks with time granularities of 1s and 5s to a certain extent; a network with a time granularity of 5s can reduce the cumulative prediction errors of a network with a time granularity of 1s to a certain extent.

[0085] Based on the same inventive concept, the present invention also proposes a regional electric power prediction system, comprising:

[0086] The acquisition module is used to acquire all power nodes in the target area containing power absorbing or power transmitting devices.

[0087] The extraction module is used to extract the spatiotemporal dynamic characteristics of each power node.

[0088] The prediction module is used to predict the power of the target area at a future time according to the spatiotemporal dynamic characteristics of each power node.

[0089] The extraction module includes:

[0090] The division unit is used to spatially grid the target area to obtain multiple grids containing only a single power node.

[0091] The electrical distance calculation unit is used to obtain the electrical distance between the power nodes with electrical association by using the equivalent impedance method, and divide the power nodes with electrical association into a power path according to the electrical distance.

[0092] The closeness centrality calculation unit is used to calculate the closeness centrality of each power node in each power path based on the electrical distance.

[0093] The allocation unit is used to allocate attention weights to power nodes according to electrical distance and proximity centrality.

[0094] The extraction unit is used to input the power node data after the attention weights are assigned into a three-dimensional convolutional network with different time granularities to obtain the spatiotemporal dynamic characteristics of the power nodes.

[0095] Based on the same inventive concept, the present invention also proposes a regional electric power prediction computer device, comprising: a memory, a processor, and a computer program stored in the memory, and the processor implements the steps of the regional electric power prediction method when executing the computer program.

[0096] Based on the same inventive concept, the present invention also proposes a readable storage medium, which stores a computer program. The computer program includes program instructions. When the program instructions are executed by a processor, they are used to execute the steps of the regional electric power prediction method.

[0097] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A method for predicting regional electric power, characterized in that: The following steps are involved: Collect all power nodes in the target area that contain power absorbing or transmitting devices; Extract the spatiotemporal dynamic characteristics of each power node; Predict the power of the target area at future times based on the spatiotemporal dynamic characteristics of each power node; The step of extracting the spatiotemporal dynamic characteristics of each power node specifically includes the following steps: The target area is spatially meshed to obtain multiple meshes containing only one power node; An equivalent impedance method is used to obtain the electrical distance between the electrically associated power nodes, and the electrically associated power nodes are divided into a power path according to the electrical distance; Calculate the closeness centrality of each power node in each power path based on electrical distance; Assign attention weights to power nodes based on electrical distance and proximity centrality; The power node data after allocating attention weights is input into a three-dimensional convolutional network with different time granularities to obtain the spatiotemporal dynamic characteristics of the power nodes; the spatiotemporal dynamic characteristics are obtained through three-dimensional convolution operations to obtain the dynamic laws of the power of the power nodes under the comprehensive influence of time, electrical distance and proximity centrality.

2. A method for regional electric power prediction according to claim 1, characterized in that: The spatial gridding of the target area to obtain a plurality of grids including only one power node specifically includes the following steps: Divide the target area into grids of different sizes according to the power density at different locations; select the largest grid to divide the area with sparse power node distribution and concentrated power density; use the smallest grid as the unit grid, and divide the other grids according to the unit grid size; The power nodes are allocated to each grid to obtain a two-dimensional grid containing power nodes; each grid contains only one power node; if there is a power node spanning multiple grids, the power node is divided into one of the grids that does not contain other power nodes according to the minimum center distance standard.

3. A method for regional electric power prediction according to claim 1, characterized in that: The method of using the equivalent impedance method to obtain the electrical distance between the electrically associated power nodes specifically includes the following steps: For power nodes in a single power network in a target area, the modulus of the equivalent impedance between the power nodes is used as the electrical distance between the power nodes; For the power nodes in the multi-power network in the target area, the multi-power network is split into several single-power networks through the downstream tracing method to obtain the impedance between the power nodes under the action of any power source; after the equivalent impedances between the corresponding nodes under the action of several power sources are connected in parallel, the modulus value is taken as the electrical distance between the power nodes.

4. A method for regional electric power prediction according to claim 1, characterized in that: The calculation formula of the closeness centrality of each power node in each power path based on the electrical distance is expressed as: Among them, d i represents the average electrical distance from node i to each point with which it has a power transmission relationship, N represents the total number of nodes on the power path of node i, and d ij represents the electrical distance between node i and node j, CC i represents the closeness centrality of node i.

5. A method for regional electric power prediction according to claim 4, characterized in that: The method of allocating attention weights to power nodes according to electrical distance and proximity centrality specifically includes the following steps: The electrical distance and proximity centrality are normalized, and the electrical distance between two power nodes on a power path after normalization is recorded as a i ; If there is a node that is electrically associated with multiple nodes, the average of the multiple electrical distances is used; The normalized value of the closeness centrality in the power path is denoted as b i ; Assign weights to electrical distance and proximity centrality; if the electrical distance weight is λ1, the proximity centrality weight is 1-λ1, and 1>λ1>0, then the attention weight of the power node is a i ·λ1+b i ·(1-λ1).

6. A regional electric power prediction system, characterized in that: include: A collection module, used to collect all power nodes in the target area containing devices that absorb power or send power; An extraction module, used to extract the spatiotemporal dynamic characteristics of each power node; A prediction module is used to predict the power of the target area at a future time according to the spatiotemporal dynamic characteristics of each power node; Wherein, the extraction module comprises: A division unit is used to spatially grid the target area to obtain multiple grids containing only one power node; An electrical distance calculation unit, used for obtaining the electrical distance between the power nodes with electrical association by using an equivalent impedance method, and dividing the power nodes with electrical association into a power path according to the electrical distance; A closeness centrality calculation unit, used to calculate the closeness centrality of each power node in each power path based on the electrical distance; an allocation unit, for allocating attention weights to power nodes according to electrical distance and proximity centrality; The extraction unit is used to input the power node data after the attention weights are assigned into a three-dimensional convolutional network with different time granularities to obtain the spatiotemporal dynamic characteristics of the power nodes; the spatiotemporal dynamic characteristics are obtained through a three-dimensional convolution operation to obtain the dynamic laws of the power of the power nodes under the comprehensive influence of time, electrical distance and proximity centrality.

7. A computer device for regional electric power prediction, characterized in that: include: A memory, a processor, and a computer program stored in the memory, wherein the processor implements the steps of the regional electric power prediction method according to any one of claims 1 to 5 when executing the computer program.

8. A readable storage medium, characterized in that: The readable storage medium stores a computer program, and the computer program includes program instructions. When the program instructions are executed by a processor, they are used to execute the steps of the regional electric power prediction method according to any one of claims 1 to 5.