Space power load prediction method based on interpretable artificial intelligence

Abstract

The invention discloses a spatial power load prediction method based on interpretable artificial intelligence, which belongs to the field of power distribution network planning in a power system, and is a spatial load prediction method based on a time sequence generative adversarial network TimeGAN, a graph attention network GAT and an iTransform. Comprising the steps that a geographic information system is established, TimeGAN is adopted to carry out data enhancement on II-type cellular loads, a condition structure in a time sequence is fully utilized, and the limitation of a traditional data generation method can be well overcome; according to the space power load prediction model based on the GAT-iTransform, the time correlation and the space correlation of cellular loads are synchronously considered, the precision and the generalization ability of the prediction model are effectively improved, interpretability of the model in the time dimension and the space dimension is achieved through an attention mechanism, and stable and reliable data support can be provided for power system planning.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of distribution network planning in power systems, and particularly relates to a spatial power load forecasting method based on explainable artificial intelligence. BACKGROUND

[0002] Spatial power load forecasting, also known as spatial load forecasting (SLF), is a key link in urban power grid planning, which aims to predict the power load demand and its geographical distribution in a target area in the future. Accurate spatial load forecasting results are directly related to the rationality and economy of power grid architecture design, substation site selection, line corridor planning and power equipment configuration, and play an important supporting role in ensuring the safe, reliable and efficient operation of power grid.

[0003] At present, with the development of artificial intelligence technology, a large number of studies have begun to use deep learning models for spatial load forecasting in order to improve the prediction accuracy. However, such models generally have the "black box" feature: the prediction mechanism is highly complex, with numerous internal parameters that are difficult to assign clear physical meanings, resulting in an opaque prediction process and results that are not easy to interpret. In the field of power grid planning, which requires high decision-making credibility, the lack of model explainability seriously hinders the practical application value of the prediction results. In addition, existing deep learning methods are highly dependent on data quality and quantity. In scenarios where historical data is incomplete, sample data is insufficient or new regional data accumulation is lacking, the model generalization ability significantly decreases, the prediction accuracy cannot be guaranteed, and even the prediction fails.

[0004] Therefore, how to maintain prediction accuracy while enhancing the explainability of spatial load forecasting models and improving their adaptability in data-scarce conditions has become a key problem to be solved in the development of current power grid planning technology. SUMMARY

[0005] The purpose of the present application is to solve the two bottlenecks in spatial load forecasting:

[0006] 1. Data scarcity problem: the prediction performance of traditional models significantly decreases under insufficient data conditions.

[0007] 2. Model limitation problem: existing models are difficult to fully explore and explain the potential interaction relationships between different spatially distributed type II cells.

[0008] Therefore, the present application proposes a spatial power load forecasting method considering the explainability of artificial intelligence models and the spatio-temporal correlation of cell load, which is scientific and reasonable, highly applicable, and can effectively improve the prediction accuracy in data-limited scenarios.

[0009] The application aims to realize the technical scheme of a space power load prediction method based on an interpretable artificial intelligence, comprising the following steps:

[0010] S1: Establishing a power geographic information system, generating I-type cells and II-type cells with different spatial resolutions based on multi-source data in a planning area;

[0011] S2: Constructing a data enhancement model, using a time series generative adversarial network to enhance the load time series data of II-type cells to generate extended data conforming to the original time series distribution thereof;

[0012] S3: Constructing a spatio-temporal information graph of II-type cells and using a graph attention network to extract spatial features of the load of each II-type cell, wherein the spatial features reflect the spatial correlation thereof by calculating the spatial attention weight between cells;

[0013] S4: Splicing the spatial features and the time series of the load of II-type cells to form joint features and inputting the joint features into an iTransformer network, using the multi-head self-attention mechanism of the iTransformer network to extract the time features of the load of II-type cells;

[0014] S5: Obtaining the space power load prediction result of the target year through the output layer of the iTransformer network based on the extracted spatio-temporal features.

[0015] Further, in step S1, the multi-source data comprises a grid structure, land use information, 10kV feeder power supply range and historical load data; the I-type cells are generated based on the power supply range of 10kV feeders, and the II-type cells are generated based on square grid division of equal size.

[0016] Further, in step S2, the data enhancement model comprises an embedding network, a reproduction network, a generator, a discriminator and a supervisor;

[0017] The embedding network is used to map the real high-dimensional static data and dynamic sequence data of II-type cells to a hidden space to obtain corresponding low-dimensional representations;

[0018] The generator is used to map Gaussian noise to the hidden space to generate simulated low-dimensional static data and dynamic sequences;

[0019] The supervisor is used to extract the time-dependent features in the real low-dimensional dynamic sequences; based on the extracted time-dependent features, the generator is constrained so that the dynamic sequences output by the generator conform to the time series characteristics;

[0020] The discriminator is used to distinguish the authenticity of the input sequences;

[0021] The reproduction network is used to restore the low-dimensional data in the latent space to the high-dimensional space, and complete data enhancement.

[0022] Further, the operation of the embedding network follows the formula:

[0023]

[0024] wherein s represents the real high-dimensional static data of the II-type cell; h s represents the low-dimensional static data of the cell after s is mapped to the latent space; x t represents the high-dimensional vector composed of the load of each cell at time t; h t represents the load of each cell at time t after x t is mapped to the latent space; h t-1 represents the low-dimensional vector composed of the load of each cell at time t-1 after x t-1 is mapped to the latent space; e(·) represents a recurrent neural network function in the embedding network, which is implemented by a gated recurrent unit.

[0025] Further, the recurrent neural network function adopted in the embedding network and the generator is a gated recurrent unit; the bidirectional recurrent neural network function adopted in the reproduction network and the discriminator is a bidirectional gated recurrent unit.

[0026] Further, in step S3, the construction of the spatiotemporal information graph of the II-type cell is specifically: taking the geometric center of the II-type cell as the node of the graph, taking the adjacency relationship between the cells as the edge of the graph, and taking the load data of the cell as the feature information of the node, thereby constructing the graph structure.

[0027] Further, in step S3, the extraction of the spatial feature by the graph attention network includes:

[0028] determining the adjacency matrix according to the adjacency of the cells in geographical position and the distance of the center coordinates of each cell;

[0029] calculating the attention coefficient between the target cell and its adjacent cells, which is obtained by a shared attention function acting on the node features after splicing and linear transformation;

[0030] normalizing the attention coefficient to obtain the spatial attention weight between the target cell and its adjacent cells;

[0031] weighting and summing the features of the adjacent cells according to the spatial attention weight, and obtaining the updated spatial feature of the target cell through an activation function.

[0032] Further, the calculation of the attention coefficient follows the formula:

[0033]

[0034] wherein e ij denotes the attention coefficient of cell i to cell j; a(·) denotes the attention function, which is used to receive the spliced high-dimensional features and map them into a scalar; || denotes the feature splicing operation; and denotes the input feature information corresponding to cell i and cell j; W denotes the shared weight matrix for linear mapping; j∈N i , N i denotes the set of cells adjacent to cell i.

[0035] Further, in step S4, the use of iTransformer to extract time characteristics specifically includes: using an independent embedding method to map the load time series of each II type cell into a feature vector, and capturing the long-term dependence in the time series through a multi-head self-attention mechanism and layer normalization operation.

[0036] Further, in step S5, the output layer of the iTransformer network is a full connection layer. A parallel one-dimensional convolutional neural network (1D-CNN) module is introduced in the full connection layer, which is used to compress and map the high-dimensional spatio-temporal features, and the outputs of the two are fused through a skip connection, so as to fully utilize the spatio-temporal features that are ignored in the mapping process of the Projection module, and improve the expression and information utilization ability of the model in the output stage. The prediction output module expression is:

[0037] M O (F)=M P (F)+M C (F)=f n +M C (F)

[0038] wherein F denotes a feature matrix; f n denotes the nth sequence feature of the feature matrix F; M O denotes the full connection layer of the prediction model; M P (F) denotes the Projection module; M C (F) denotes a one-dimensional convolutional neural network (1D-CNN) module.

[0039] Compared with existing technologies, this invention proposes a spatial load prediction method based on time-series generative adversarial networks (TimeGAN), graph attention networks (GAT), and iTransformer. TimeGAN is used to augment the data of type II cellular loads, fully utilizing the conditional structure in the time series, which effectively overcomes the limitations of traditional data generation methods. The spatial power load prediction model based on GAT-iTransformer simultaneously considers the temporal and spatial correlations of cellular loads, effectively improving the accuracy and generalization ability of the prediction model. Furthermore, the attention mechanism enables the model to be interpretable in both time and space dimensions, providing stable and reliable data support for power system planning. Attached Figure Description

[0040] Figure 1 This is a basic principle diagram of a spatial power load forecasting method based on interpretable artificial intelligence.

[0041] Figure 2 This is a land use information map of the area to be predicted;

[0042] Figure 3 Architecture diagram for augmenting type II cell load data;

[0043] Figure 4 The structure diagram of the spatial load prediction model that takes into account the spatiotemporal correlation between cells;

[0044] Figure 5 A graph showing the measured values ​​for the target year and the predicted values ​​for each method;

[0045] Figure 6 Heatmap of temporal and spatial attention weights. Detailed Implementation

[0046] To make the objectives, features, and advantages of this invention more apparent and understandable, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, this invention is not limited to the following embodiments, and specific implementation methods can be determined according to the technical solutions of this invention and actual circumstances. To avoid obscuring the essence of this invention, well-known methods, processes, and procedures are not described in detail.

[0047] A spatial power load forecasting method based on interpretable artificial intelligence includes establishing a geographic information system (GIS), constructing a data augmentation model, extracting spatial features of type II cells using GAT, extracting temporal features of type II cells using iTransformer, and establishing a spatial power load forecasting model based on interpretable artificial intelligence.

[0048] The specific content is as follows:

[0049] S1: Establish a geographic information system based on the desktop geographic information system software MapInfo;

[0050] S11: Collect and integrate the grid structure, land use information, 10kV feeder power supply range and corresponding historical load data within the planning area; this step is existing technology and will not be described in detail here.

[0051] S12: In the power geographic information system (GIS) environment, construct a registration base map of the planning area; in the registration base map, according to the regional planning scheme, divide the land use type of the area to be predicted into eight categories: residential area, commercial area, cultural and entertainment area, industrial area, administrative office area, green space, municipal facilities area and special land, construct a land use information layer, and then establish Class I and Class II cell layers according to the power supply range of 10kV feeder and gridding technology.

[0052] S2: Build a data augmentation model;

[0053] TimeGAN is used to generate a sufficient amount of data that conforms to the temporal distribution of type II cell loading. The calculation steps are as follows:

[0054] S21: The principle of embedded networks is shown in formula (1):

[0055]

[0056] Where s represents the true high-dimensional static data of type II cells; h s This represents the low-dimensional static data of the cell after mapping s to the latent space; x t h represents the high-dimensional vector formed by the cell loads at time t; t Indicates x t The low-dimensional vector formed by the cell loadings at time t after mapping to the latent space; h t-1 Indicates x t-1 The low-dimensional vector formed by the cell loads at time t-1 after mapping to the latent space; e(·) represents the recurrent neural network function embedded in the network, and the recurrent neural network function used is a gated recurrent unit.

[0057] S22: The principle of reproducing the network is shown in formula (2):

[0058]

[0059] in, Indicates h sMapping from the latent space to high-dimensional static data of cells in the reproducible network; Indicates h t The high-dimensional vector formed by mapping the latent space to the cell loads at time t in the reproducible network; r(·) represents the bidirectional recurrent neural network function in the reproducible network, and the recurrent neural network function used is a bidirectional gated recurrent unit.

[0060] S23: The principle of the generator is shown in formula (3):

[0061]

[0062] in, This represents the low-dimensional static data of the simulated cells; g(·) represents the recurrent neural network function of the generator, which is a gated recurrent unit; z s With z t These are random samples belonging to the static and dynamic vector spaces, respectively. This represents the low-dimensional vector of the cell load at time t-1 generated by the simulation.

[0063] S24: The principle of the discriminator is shown in formula (4):

[0064]

[0065] in, This indicates that the classification results are generated from low-dimensional static data of type II cells; This represents the classification result of the low-dimensional vector generated by the loads of each cell at time t. d(·) represents the joint output of the generator and discriminator, and d(·) represents the classification discriminant function, which is implemented by a bidirectional recurrent neural network with a feedforward output layer. The recurrent neural network function used is a bidirectional gated recurrent unit.

[0066] S3: Extracting spatial features of type II cells using GAT;

[0067] S31: Construct a cellular load spatiotemporal information graph. Multi-source data, including 10kV feeder power supply range, land use information, spatial coordinates, and historical load data, are transformed into spatial loads characterized by type II cells using spatial power gridding technology. The geometric center of the type II cells is used as the node of the graph, the adjacency relationship of the cells is used as the edge of the graph, and the cellular load data is used as the feature information of the nodes to construct the cellular spatiotemporal information graph.

[0068] S32: Construct the adjacency matrix of cells, and determine the adjacency matrix based on the geographical proximity of cells and the distance between the center coordinates of each cell;

[0069] S33: The principle of the attention coefficient is shown in formula (5):

[0070]

[0071] Among them, e ij represents the attention coefficient of cell i to cell j (cell j is a neighboring cell of cell i); a(·) represents the attention function, which receives the concatenated high-dimensional features and maps them to a scalar; ‖ represents the feature concatenation operation; and Represents the input feature information corresponding to cells i and j; W represents the shared weight matrix used for linear mapping; j∈N i N i This represents the set of cells adjacent to cell i.

[0072] S34: The principle of spatial attention weights is shown in formula (6):

[0073]

[0074] Where LeakyReLU(·) represents the nonlinear transformation function; softmax(·) represents the normalization operation; softmax j (e ij ) represents the attention coefficient e of cell i to its j-th neighboring cell. ij Perform normalization; α ij This represents the spatial attention weight of cell i to cell j.

[0075] S35: The principle of spatial characteristics is shown in formula (7):

[0076]

[0077] Where σ(·) represents the sigmoid activation function; α ij α represents the spatial attention weight of cell i to cell j, i.e., the degree of influence of cell j on cell i; ii This represents the self-attention weight of cell i; This represents the spatial characteristics of cell i.

[0078] S4: Extracting the temporal features of type II cells using iTransformer;

[0079] S41: The principle of multi-head self-attention is shown in formula (8):

[0080]

[0081] Where H represents the extracted time feature matrix; Q represents the query vector; K represents the key vector, K T V represents the transpose of the key vector; V represents the value vector; d KThis represents the feature dimension of the key vector K.

[0082] S42: The principle of layer normalization (LayerNorm) is shown in formula (9):

[0083]

[0084] Among them, h l E(·) represents the eigenvector of the l-th sequence to be normalized in the time feature matrix H, where l = 1, 2, ..., L; E(·) represents the mean function, used to calculate the average value of the input sequence; D(·) represents the variance function.

[0085] S5: Establish a spatial power load forecasting model based on interpretable artificial intelligence

[0086] First, a spatiotemporal information map of type II cells is constructed. Second, spatial attention weights of adjacent cells are calculated using GAT and the spatial features of the target cell are updated. The extracted spatial features are concatenated with the time series of the target cell's load and fed into the iTransformer network as joint input. Then, the temporal features of the type II cell load are extracted using an independent embedding method. Finally, the spatial power load prediction results for the target year are output through the output layer of the iTransformer, thereby achieving joint prediction of the magnitude and spatial distribution of future spatial power load.

[0087] This invention provides a spatial power load forecasting method based on interpretable artificial intelligence. It is a spatial load forecasting method based on TimeGAN and GAT-iTransformer, and its basic principle diagram is shown below. Figure 1 As shown, the examples of the method include: basic data integration, establishment of a power geographic information system (GIS), generation of type II cellular load data, type II cellular load data enhancement model, SLF model considering the spatiotemporal correlation of cells, spatial load forecast (SLF) results and error assessment.

[0088] Based on the SLF method proposed in this invention, taking a portion of the administrative region of a city in Northeast my country as an example, historical load data from 2021 to 2023 are used to perform data augmentation and predict the Type II cell load value for this region in 2024. To verify the accuracy of the method proposed in this invention, the prediction results are compared and analyzed with those of traditional methods such as GRU, Transformer, GCN-LSTM, and STGAT. The specific details are as follows:

[0089] 1) Establish a power GIS (Geographic Information System)

[0090] A portion of the administrative districts of a city in Northeast my country was selected as the area to be predicted. Historical load data from 2021 to 2023, with a sampling interval of 5 minutes, were used to predict the spatial power load in 2024.

[0091] First, a GIS of the area to be predicted is established, and then the regional planning scheme is followed. Figure 2 The land use types of the area to be predicted are divided into eight categories: residential area (yellow), commercial area (red), cultural and entertainment area (orange), industrial area (brown), administrative office area (purple), green space (green), municipal facilities area (blue), and special-use land (dark green). Then, Class I and Class II cell layers are created according to the power supply range of the 10kV feeder and gridding technology. It should be noted that in the environment of the power geographic information system, two types of cells are generated according to different scales of power load spatial resolution: Class I cells are generated by dividing the area to be predicted according to the power supply range of the 10kV feeder, and Class II cells are generated by dividing the area according to square grids of equal size. Based on this, spatial load prediction for the area in 2024 is then performed.

[0092] 2) Type II Cell Load Data Augmentation Model

[0093] This invention utilizes TimeGAN to generate 720 expanded samples based on 36 months of measured data from 2021 to 2023. According to the correlation between the generated data and the corresponding monthly load, the data is divided into monthly datasets, with 20 samples per month. The data augmentation model architecture is as follows: Figure 3 As shown, compared with traditional data augmentation methods, this invention introduces supervised training and an autoencoder network.

[0094] First, the basic data of all type II cells are input into the embedding network of the autoencoder network for dimensionality reduction and then sent to the latent space. Second, the generator maps Gaussian noise into the latent space to simulate the low-dimensional dynamic sequence of the generated cells. Simultaneously, the supervisor in the latent space learns the temporal dependency of the real low-dimensional dynamic sequence of the cells and guides the generator to generate dynamic sequences with reasonable temporal characteristics based on the supervisory loss. Then, the low-dimensional dynamic sequence of the cells generated by the generator and supervisor and the real low-dimensional dynamic sequence of the cells are input into the discriminator to distinguish between true and false sequences. Finally, the low-dimensional dynamic sequence of the cells output by the generator and the low-dimensional static data of the cells in the latent space are restored to the same dimension as the basic data of type II cells through the reproduction network in the autoencoder network, and the noise at the original zero position is suppressed by the sparse masking mechanism, thereby realizing the data augmentation of type II cell load.

[0095] 3) SLF model considering cellular spatiotemporal correlation

[0096] This invention's SLF model, building upon traditional time-series prediction, further considers the spatial dependencies between cells to improve prediction accuracy and utilizes temporal and spatial attention mechanisms to enhance the interpretability of the prediction model. First, a spatiotemporal information map of type II cells is constructed. Second, GAT is used to calculate the spatial attention weights of cell i on cell j and update the spatial features of the target cell. The extracted spatial features are concatenated with the time series of the cell load and fed as joint input into the iTransformer network. Then, independent embedding is used to avoid the mixing of features from different type II cells in the temporal dimension. A multi-head attention layer is used to capture temporal features in the time series, and a forward propagation layer is used to further enhance the expressive power of temporal features. Layer normalization addresses the issue of different numerical scales among the temporal features of type II cells. Finally, a parallel one-dimensional convolutional neural network (1D-CNN) module is introduced into the fully connected layer (projection) of the iTransformer. Skip connections enhance the model's ability to express spatiotemporal features and utilize information in the output stage, outputting the SLF result. Its principle and structure are as follows: Figure 4 As shown in the figure, this is a spatial load prediction model that takes into account the spatiotemporal correlation between cells (a spatial power load prediction model based on GAT-iTransformer), used to process the spatiotemporal information of type II cell loads. The specific explanation is as follows:

[0097] 1. Feature normalization

[0098] formula:

[0099] Specifically, the input feature x is standardized to obtain... Setting the mean μ to 0 and the standard deviation σ to 1 helps stabilize the model during training.

[0100] 2. Layer Normalization (LayerNorm) is used to normalize the output of a specific layer in a neural network.

[0101] 3. Tokens and Embedding

[0102] Token: Represents the basic unit (such as time step or cell) in the input sequence.

[0103] Embedded Token: Maps the token to a dense vector representation.

[0104] 4. Constructing a spatiotemporal information graph of type II cellular loads

[0105] The figure shows multiple cells (C1 to C9) and their spatial distribution, forming a graph structure.

[0106] Each cell represents a spatial node, and edges represent spatial relationships.

[0107] 5. Graph Attention Network (GAT) for extracting spatial features

[0108] GAT: Used to extract spatial features of nodes in a graph.

[0109] step:

[0110] Calculate the attention coefficient e ij ;

[0111] The spatial attention weight α of cell i to cell j is obtained using softmax normalization. ij ;

[0112] Update node characteristics.

[0113] 6. iTransformer section

[0114] Multi-head attention layer: processes sequential data and captures dependencies.

[0115] Forward propagation layer: A fully connected layer that performs nonlinear transformations.

[0116] Fully connected layer (projection): Maps features to the target dimension.

[0117] 7. A one-dimensional convolutional neural network module is introduced in parallel in the fully connected layer to compress and map high-dimensional spatiotemporal features.

[0118] It should be noted that feature normalization, layer normalization, token embedding, graph attention network (GAT), multi-head attention mechanism, forward propagation layer, softmax, 1D-CNN and fully connected layer (projection) are existing technologies. The core innovation of this model is to propose a hybrid architecture of graph attention network and iTransformer to solve the spatiotemporal prediction problem of type II cell load.

[0119] Although the model is built upon mature components such as feature normalization, layer normalization, and multi-head attention, its innovation lies in the collaborative modeling mechanism of spatiotemporal dependencies:

[0120] Spatial Dependency Capture: By introducing a graph attention network, distributed type II cells are abstracted as nodes in a graph structure. GAT accurately captures the complex spatial nonlinear effects caused by factors such as geographical location and similarity of electricity consumption characteristics by adaptively learning the association weights between nodes (i.e., between cells).

[0121] Temporal dynamic modeling: Subsequently, the time series of each cell is deeply processed by the iTransformer encoder. Its multi-head self-attention mechanism can effectively capture the long-range, multi-scale temporal dependencies in the load data and identify the importance of different time points for future prediction.

[0122] Ultimately, this model achieves a unified modeling of the core characteristic of Type II cellular loads—"interrelated in space and dynamically evolving in time"—through end-to-end fusion of GAT (spatial) and iTransformer (temporal), significantly improving prediction accuracy.

[0123] 4) SLF Results and Error Assessment

[0124] The maximum value of the load for each type II cell in the prediction results is taken to form the SLF result for type II cells in the target year. To demonstrate the accuracy of the SLF method of this invention, its prediction results are compared with those of the GRU method, Transformer, GCN-LSTM, STGCN, STGAT method, and GAT-Transformer method. The SLF prediction results and their prediction amplitude errors are shown in [reference needed]. Figure 5 As shown in Table 1.

[0125] Table 1. Amplitude error of predicted Class II cell load values

[0126]

[0127]

[0128] In Table 1, compared with the GRU prediction method and the Transformer prediction method, the MAE (mean absolute error) of the method of this invention is reduced by 36.2 kW and 33.2 kW, respectively, and the MAPE (mean relative error) is reduced by 15.11% and 11.44%, respectively. Both of these methods only consider the temporal development law of the cellular load itself, while the present invention comprehensively considers the spatial and temporal effects of the cellular load, thus achieving higher prediction accuracy. This demonstrates the importance of the spatial distribution characteristics of the load to the prediction accuracy of SLF.

[0129] Furthermore, the MAE and MAPE of the GRU and Transformer forecasting methods are higher than those of the GCN-LSTM and STGAT forecasting methods. This further demonstrates that considering only the temporal development of cellular loads is insufficient when forecasting spatial loads. The forecasting results of all methods are then visualized to generate spatial load maps for comparison. Figure 5 As shown.

[0130] In the spatial dimension, the interpretability of the predictive model's structure-aware capability is achieved using a spatial attention mechanism. This is illustrated using graph data samples from three different time steps as an example. Figure 6 As shown. By Figure 6 (a) shows that the prediction model no longer focuses on all cells equally, but instead assigns higher attention to "high-information" regions based on the differences in features between cells, identifying some representative key cell regions, demonstrating a differentiated perception strategy and spatial focusing ability; in the time dimension, taking the temporal attention heatmap of one cell as an example... Figure 6 (b) shows the heatmap changes of the attention weights of cell i to the time step in the iTransformer model at iterations 1, 50, and 100, reflecting the evolution of the model's attention pattern to the time dimension during training. At iteration 1 ( Figure 6 (b) Above), the attention distribution is relatively scattered and lacks a clear pattern, indicating that the model has not yet effectively captured the potential dependencies in the time series. As training progresses to the 50th iteration ( Figure 6 (b) In the heatmap, a distinct striped structure begins to appear, indicating that the model has initially learned the correlation between different time steps and is able to focus its attention on time segments with higher information content. When the iteration reaches the 100th time ( Figure 6 (b) below, the heatmap presents a clear and regular structural pattern, demonstrating that the model has been able to accurately identify the key temporal information most valuable to the prediction task. This gradually evolving attention pattern not only reflects the effectiveness of iTransformer in temporal modeling but also enhances the interpretability of the model in time series prediction tasks.

[0131] 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 space power load prediction method based on interpretable artificial intelligence, characterized by, The method comprises the following steps: S1: establishing a power geographic information system, generating I-type cells and II-type cells with different spatial resolutions based on multi-source data in the planning area; S2: constructing a data enhancement model, using a time series generative adversarial network to enhance the load time series data of the II-type cells to generate extended data conforming to the original time series distribution thereof; S3: constructing a spatio-temporal information graph of the II-type cells and using a graph attention network to extract spatial features of the load of each II-type cell, wherein the spatial features reflect the spatial correlation between the cells by calculating the spatial attention weights between the cells; S4: splicing the spatial features and the time series of the load of the II-type cells to form joint features, and inputting the joint features into an iTransformer network to extract time features of the load of the II-type cells by using the multi-head self-attention mechanism of the iTransformer network; S5: obtaining the spatial power load prediction result of the target year based on the extracted spatio-temporal features through the output layer of the iTransformer network. 2.The spatial power load forecasting method based on explainable artificial intelligence according to claim 1, wherein, In step S1, the multi-source data comprises a grid structure, land use information, 10kV feeder power supply range and historical load data; the I-type cells are generated based on the power supply range of the 10kV feeder; and the II-type cells are generated based on square grid division of equal size. 3.The spatial power load forecasting method based on explainable artificial intelligence according to claim 1, wherein, In step S2, the data enhancement model comprises an embedding network, a reproduction network, a generator, a discriminator and a supervisor; The embedding network is used to map the real high-dimensional static data and dynamic sequence data of the II-type cells to a hidden space to obtain corresponding low-dimensional representations; The generator is used to map Gaussian noise to the hidden space to generate simulated low-dimensional static data and dynamic sequences; The supervisor is used to extract time-dependent features in the real low-dimensional dynamic sequences; Based on the extracted time-dependent features, the generator is constrained so that the dynamic sequences output by the generator conform to the time series characteristics; The discriminator is used to distinguish the authenticity of the input sequences; The reproduction network is used to restore the low-dimensional data in the hidden space to the high-dimensional space to complete data enhancement. 4.The spatial power load forecasting method based on explainable artificial intelligence according to claim 3, wherein, The operation of the embedding network follows the following formula: where s represents the real high-dimensional static data of the type II cell; h s represents the low-dimensional static data of the cell after s is mapped to the hidden space; x t represents the high-dimensional vector composed of the load of each cell at time t; h t represents the low-dimensional vector composed of the load of each cell at time t after x t is mapped to the hidden space; h t-1 represents the low-dimensional vector composed of the load of each cell at time t-1 after x t-1 is mapped to the hidden space; e(·) represents a recurrent neural network function in the embedding network, which is implemented by a gated recurrent unit. 5.The spatial power load forecasting method based on explainable artificial intelligence according to claim 3, wherein, The recurrent neural network function adopted in the embedding network and the generator is a gated recurrent unit; and the bidirectional recurrent neural network function adopted in the reproduction network and the discriminator is a bidirectional gated recurrent unit. 6.The spatial power load forecasting method based on explainable artificial intelligence according to claim 1, wherein, In step S3, the construction of the spatio-temporal information graph of the II-type cells specifically comprises: taking the geometric center of the II-type cell as a node of the graph, taking the adjacency relationship between the cells as an edge of the graph, and taking the load data of the cell as the feature information of the node, so as to construct the graph structure. 7.The spatial power load forecasting method based on explainable artificial intelligence according to claim 6, wherein, In step S3, the extraction of the spatial features by the graph attention network comprises: determining an adjacency matrix according to the adjacency of the cells in geographical position and the distance between the center coordinates of the cells; calculating the attention coefficients between the target cell and its adjacent cells, wherein the attention coefficients are obtained by acting on the node features after splicing and linear transformation through a shared attention function; normalizing the attention coefficients to obtain the spatial attention weights between the target cell and its adjacent cells; and According to the spatial attention weight, features of adjacent cells are weighted and summed, and an updated spatial feature of the target cell is obtained through an activation function. 8.The spatial power load forecasting method based on explainable artificial intelligence according to claim 7, wherein, The calculation of the attention coefficient follows the formula: where e ij denotes the attention coefficient of cell i to cell j; a(·) denotes the attention function, which is used to receive the spliced high-dimensional features and map them into a scalar; || denotes the feature splicing operation; and denotes the input feature information corresponding to cell i and cell j; W denotes a shared weight matrix for linear mapping; j∈N i , N i denotes the set of cells adjacent to cell i. 9.The spatial power load forecasting method based on explainable artificial intelligence according to claim 1, wherein, In step S4, the time feature extraction using iTransformer specifically includes: mapping the load time series of each II type cell into a feature vector in an independent embedding manner, and capturing long-term dependence in the time series through a multi-head self-attention mechanism and layer normalization operation. 10.The spatial power load forecasting method based on explainable artificial intelligence according to claim 1, wherein, In step S5, the output layer of the iTransformer network is a full connection layer, a one-dimensional convolutional neural network module is introduced in parallel in the full connection layer, and the outputs of the two are fused through a skip connection.

Images