Subway power supply system load prediction method and device based on space-time diagram neural network

By building a spatiotemporal graph neural network, combining geographical distance graphs and load similarity graphs, and dynamically learning the graph structure, the problem that traditional methods are difficult to capture the space-time characteristics of the load of subway power supply systems is solved, and accurate load prediction and adaptability improvement is achieved. Subway operation and management provide guarantees.

CN118970943APending Publication Date: 2024-11-15NR ELECTRIC CO LTD +1
View PDF 0 Cites 3 Cited by

Patent Information

Application Number
CN202411142497.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-20
Publication Date
2024-11-15

AI Technical Summary

Technical Problem

The traditional load prediction method of subway power supply systems is difficult to effectively capture the spatial and temporal characteristics of load, especially in the face of complex characteristics such as load mobility, time deformation and nonlinearity, and the long sequence diagram structure cannot reflect the real connection relationship.

Method used

Using a method based on spatiotemporal graph neural network, the geographic distance map and load similarity map are constructed, combined with the DynGraph dynamic layer and mixhop graph network propagation layer, dynamically learn the graph structure and parameters, capture the spatiotemporal features in the load data, and realize load prediction. .

Benefits of technology

Accurate prediction of the load of the subway power supply system is achieved, the complexity of the model is reduced, the number of parameters and the risk of overfitting is reduced, the adaptability of the model is improved, and the subway operation and management is guaranteed.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118970943A_ABST
    Figure CN118970943A_ABST
Patent Text Reader

Abstract

The invention discloses a metro power supply system load prediction method and device based on a space-time diagram neural network, and belongs to the technical field of metro power supply system load prediction, and the method comprises the steps: obtaining a space relation between metro networks and historical load data; and training a pre-constructed load prediction model through the spatial relationship and historical load data to obtain a trained load prediction model, the pre-constructed load prediction model continuously updates learnable parameters in the load prediction model while performing dynamic learning according to the historical load data; the time sequence information needing to be predicted is input into the trained load prediction model, and a load prediction result is obtained.According to the load prediction method, the spatial relation between subway networks is seen from multiple perspectives, the spatial-temporal dynamic change of the load of the subway power supply system is comprehensively captured, and the accuracy of load prediction of the subway power supply system is greatly improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of load prediction of subway power supply systems, and in particular relates to a method and device for load prediction of subway power supply systems based on spatiotemporal graph neural networks. Background Art

[0002] Time series forecasting enables decision-making by modeling the evolution of future states or events, and plays an important role in many fields, such as traffic, healthcare, and finance. As the core supporting facility for urban traffic operation, load forecasting of subway power supply system has always been an important issue in operation management. Traditional load forecasting methods often use models based on statistics or traditional machine learning algorithms, such as linear regression, support vector machine, etc. However, these methods often find it difficult to effectively capture the spatiotemporal characteristics of subway power supply system loads, especially in the face of complex characteristics such as load mobility, time deformation, and nonlinearity.

[0003] With the development of deep learning technology, neural networks have made significant progress in the field of time series prediction. Among them, the use of graph neural networks for multivariate time series prediction has great application prospects. As a deep learning model that combines spatiotemporal information, graph neural networks have unique advantages in processing spatiotemporal data. By modeling the topological structure and operation data of the subway power supply system, the spatiotemporal graph neural network can effectively capture the change law of load in time and space, thereby achieving accurate prediction of the load. However, the long sequence graph structure will evolve over time, and the fixed graph structure cannot reflect the real connection relationship. Therefore, the scheme of using graph neural networks to predict the load of the subway power supply system based on the existing technology still needs to be improved. Summary of the invention

[0004] The present invention provides a method and device for load prediction of a subway power supply system based on a spatiotemporal graph neural network. A pre-constructed prediction model is trained by combining the spatial information of the spatiotemporal subway network and historical load information. During the training process, a dynamic learning graph is continuously generated to capture the hidden relationship between the historical load information and update the model parameters, thereby achieving accurate load prediction through the dynamic spatiotemporal characteristics in the load data.

[0005] In order to solve the above technical problems, the present invention is implemented by adopting the following technical solutions:

[0006] In a first aspect, the present invention provides a method for predicting load of a subway power supply system based on a spatiotemporal graph neural network, comprising:

[0007] Obtain spatial relationships and historical load data between subway networks;

[0008] The pre-built load forecasting model is trained by using the spatial relationship and the historical load data to obtain a trained load forecasting model, wherein the pre-built load forecasting model dynamically learns according to the historical load data while continuously updating the learnable parameters in the load forecasting model;

[0009] The time series information to be predicted is input into a trained load prediction model to obtain a load prediction result, wherein the time series information to be predicted includes time information and site information.

[0010] Optionally, the pre-built load forecasting model includes multiple layers of spatiotemporal layers with residual connections and a prediction output module connected to the multiple layers of spatiotemporal layers, and the output of each spatiotemporal layer is connected through a fully connected layer, wherein each spatiotemporal layer includes a temporal convolution module, a dynamic learning graph generation module and a spatial graph convolution module; the temporal convolution module includes two expanded temporal convolution layers, one of which is connected to a tanh function and the other is connected to a sigmoid function; the dynamic learning graph generation module includes a DynGraph dynamic layer layer, and the spatial graph convolution module includes a mixhop graph network propagation layer.

[0011] Optionally, it is characterized in that the training of a pre-built load forecasting model according to the spatial relationship and historical load data comprises:

[0012] Construct a geographic distance map based on the geographic relationships between subway networks;

[0013] Construct a load similarity graph based on historical load data;

[0014] The geographic distance map, the load similarity map and the historical time series information are input into a pre-built load forecasting model to obtain a trained load forecasting model.

[0015] Optionally, constructing a geographical distance map according to the geographical relationship between subway networks includes:

[0016] The Euclidean distance between two sites is calculated based on the latitude and longitude of each site to obtain the geographic distance matrix D geo , define the following weighted formula to obtain the geographic distance map A geo :

[0017] ,

[0018] in, represents the distance graph between sites i and j, represents the distance matrix between sites i and j, represents the standard deviation of the distance, Indicates the set threshold.

[0019] Optionally, constructing a load similarity graph according to historical load data includes:

[0020] Arrange the time series X and Y into an n×m grid, where each point (i, j) in the grid is X i and Y j Alignment between, value is X i and Y j The absolute value of the distance difference between them, (i k , j k ) is calculated using a dynamic programming algorithm:

[0021] ,

[0022] Where d is the Euclidean distance and the total path length is calculated using the following formula: .

[0023] Optionally, inputting the time series information to be predicted into a trained load forecasting model to obtain a load forecasting result includes:

[0024] The time convolution module in the current space-time layer performs convolution processing on the space-time graph output by the previous space-time layer and the initial time series of the input model to obtain the time series after time expansion; the dynamic learning graph generation module in the current space-time layer processes the time series after time expansion to obtain the phased graph structure information; the spatial graph convolution module in the current space-time layer performs convolution processing on the graph structure information and the time series after time expansion to obtain the space-time graph output by the current space-time layer; the space-time graph output by the current space-time layer and the initial time series of the input model are added to obtain the input data of the next space-time layer, wherein the input data of the first space-time layer is the initial time series of the input model;

[0025] According to the above processing steps, the time-space diagrams output by all time-space layers in the load forecasting model are obtained in sequence;

[0026] The prediction output module connects all spatiotemporal graphs through a fully connected layer to obtain the predicted output sequence.

[0027] Optionally, the convolution processing of the spatiotemporal graph output by the previous spatiotemporal layer and the time series of the input model by the temporal convolution module in the current spatiotemporal layer includes:

[0028] The tanh function connected after the first temporal convolution layer in the temporal convolution module is used as a filter, and the sigmoid function connected after the second temporal convolution layer is used as a gate controller. The output of the temporal convolution module is:

[0029] ,

[0030] Among them, T1 represents the time series after convolution output by the first time convolution layer, T2 represents the time series after convolution output by the second time convolution layer, and T out Represents the convolved time series output by the temporal convolution module.

[0031] Optionally, the dynamic learning graph generation module in the current spatiotemporal layer processes the time series after time expansion to obtain phased graph structure information, including:

[0032] Cut the historical time series and extract features from the cut historical time series;

[0033] Perform sliding average on the time series after time expansion, and add the time series after sliding average to the extracted features;

[0034] The summed time series is dynamically learned through the DynGraph dynamic layer. The learning process is as follows:

[0035] ,

[0036] Among them, X t Indicates the trend characteristics of the sequence, X s represents the seasonal characteristics of the sequence, X w represents the weather characteristics of the sequence, AvgPool represents the sliding average operation, and α represents the sum of the trend sequence, seasonal sequence, and weather sequence after the linear layer;

[0037] Two linear layers are used to encode the time series output by the DynGraph dynamic layer into the hidden layer to generate an adjacency matrix, where one of the two linear layers is connected to a sigmoid function as a mask. The outputs of the two linear layers and the dynamic learning module are as follows:

[0038] ,

[0039] in, and Represent the value of the learned graph structure and the value of the mask at the i-th row and j-th column, respectively. represents the sigmoid activation function, Represents the dynamic learning graph matrix of the mth time period.

[0040] Optionally, the convolution processing is performed on the graph structure information and the time series after time expansion by the spatial graph convolution module in the current spatiotemporal layer to obtain the spatiotemporal graph output by the current spatiotemporal layer, including:

[0041] The adjacency matrix output by the dynamic learning module is propagated through the mixhop graph network propagation layer, wherein the information propagation has k steps, and the calculation process of each step is as follows:

[0042] ,

[0043] in, is a hyperparameter used to control the retention rate of the original state information. represents the state of the previous hidden layer, Represents the adjacency matrix generated by the dynamic learning module;

[0044] Information selection is performed on the k-step information propagation results through the mixhop graph network propagation layer. The calculation process of the information selection is as follows:

[0045] ,

[0046] in, represents the hidden state of the current layer output, represents the weight matrix, .

[0047] Optionally, the prediction output module connects all spatiotemporal graphs through a fully connected layer to obtain a predicted output sequence, including:

[0048] The time series data processed by the time convolution module is convolved through four filters of different sizes, where the sizes of the convolution kernels of each filter are 1×2, 1×3, 1×6 and 1×7 respectively;

[0049] The output of each filter is connected across channels after the maximum pooling operation, and the model output is obtained as follows:

[0050] ,

[0051] Among them, z' represents the time series of the load forecasting model output, z represents the time series of the input maximum pooling layer, and f 1×k represents convolution kernels of different sizes, * represents convolution operation, max (·) represents maximum pooling operation, and concat (·) represents concatenation operation.

[0052] In a second aspect, the present invention provides a subway power supply system load prediction device based on a spatiotemporal graph neural network, comprising:

[0053] Data acquisition module: used to obtain the spatial relationship between subway networks and historical load data;

[0054] A load forecasting model training module is used to train a pre-built load forecasting model through the spatial relationship and the historical load data to obtain a trained load forecasting model, wherein the pre-built load forecasting model dynamically learns according to the historical load data and continuously updates the learnable parameters in the load forecasting model;

[0055] Load forecasting result acquisition module: used to input the time series information to be predicted into the trained load forecasting model to obtain the load forecasting result, wherein the time series information to be predicted includes time information and site information.

[0056] Compared with the prior art, the present invention achieves the following beneficial effects: the present invention views the spatial relationship between subway networks from multiple perspectives, constructs multiple graphs to represent the relationship between each station, thereby allowing the graph neural network to consider the interaction between subway networks from different perspectives, and continuously updates the learnable parameters in the load forecasting model through dynamic learning of historical time series information. Accurate load forecasting can be achieved through the dynamic spatiotemporal characteristics in the load data of the subway power supply system, providing a guarantee for efficient management and resource allocation of subway operations; residual connections are added between each spatiotemporal layer of the graph neural network to facilitate training and avoid gradient disappearance; by adopting two dilated time convolution layers, it is possible to process long time series and reduce the complexity of the model; the method of learning historical series through the DynGraph layer to obtain a dynamic learning graph captures time series correlations of different scales, introduces a spatial graph module, and propagates information through an adjacency matrix, which can better capture contextual information, reduce the number of parameters and the risk of overfitting, and improve the adaptability of the model through weighted aggregation. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 The figure is a flow chart of a method for predicting load of a subway power supply system in one embodiment of the present invention;

[0058] Figure 2 Shown is a schematic diagram of the structure of a load forecasting model in an embodiment of the present invention;

[0059] Figure 3 Shown is a schematic structural diagram of the DynGraph layer in an embodiment of the present invention. DETAILED DESCRIPTION

[0060] The present invention will be further described below in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and cannot be used to limit the protection scope of the present invention.

[0061] Example 1

[0062] like Figure 1As shown, this embodiment provides a subway power supply system load prediction method based on spatiotemporal graph neural network, including:

[0063] Obtain spatial relationships and historical load data between subway networks;

[0064] The pre-built load forecasting model is trained by using the spatial relationship and the historical load data to obtain a trained load forecasting model, wherein the pre-built load forecasting model dynamically learns according to the historical load data while continuously updating the learnable parameters in the load forecasting model;

[0065] The time series information to be predicted is input into a trained load prediction model to obtain a load prediction result, wherein the time series information to be predicted includes time information and site information.

[0066] Example 2

[0067] Based on Example 1, this example also makes the following design.

[0068] like Figure 2 As shown in the figure, the pre-built load forecasting model is composed of L layers of spatiotemporal layers stacked together, each of which includes a temporal convolution module, a dynamic learning graph generator (DGG), and a spatial graph convolution module; the temporal convolution module consists of two dilated temporal convolution layers, each of which is followed by an activation function, a tanh function as a filter, and a sigmoid function as a gate controller; the purpose of this design is to increase the receptive field of the model, thereby processing longer time series while reducing the complexity of the model; the graph convolution module includes a mixhop propagation layer, which is divided into two steps: information transmission and information selection.

[0069] The load forecasting method is divided into the following steps.

[0070] Step 1. Construct a spatiotemporal graph: View the spatial relationship between subway networks from multiple perspectives, and construct multiple graphs to represent the relationship between each station. The composition includes geographic distance graph, semantic distance graph and dynamic learning graph, which allows the graph neural network to consider the interaction between subway networks from different perspectives. Accurate load forecasting can be achieved through the dynamic spatiotemporal characteristics in the load data of the subway power supply system.

[0071] Step 1.1: Construct geographic distance map A geo

[0072] The Euclidean distance between two sites is calculated based on the latitude and longitude of each site to obtain the geographic distance matrix D geoAccording to research, the closer the distance between stations, the more similar their geographical locations and environments are, and the greater the impact on the load situation. Therefore, based on the geographical relationship of the subway network, the following weighted formula is defined to construct the geographical distance map A geo :

[0073] ,

[0074] in, represents the distance graph between sites i and j, represents the distance matrix between sites i and j, represents the standard deviation of the distance, Represents the set threshold, which is set to 0.8 here. By setting the threshold, the sparsity of the adjacency matrix is ​​increased to reduce the complexity.

[0075] Step 1.2: Construct load similarity graph A dtw

[0076] By calculating the dynamic time warping (DTW) similarity between historical load data, the principle is to find the best match between two time series by distorting one time series relative to the other. Distortion is done by stretching or compressing the time axis of a series. After the distortion operation, the distance between the two series can be minimized. The specific process is to first arrange the time series X and the time series Y into an n×m grid, and each point (i, j) in the grid is X i and Y j Alignment between, value is X i and Y j The absolute value of the distance difference between them, (i k , j k ) can be calculated using a dynamic programming algorithm:

[0077] ,

[0078] Where d is the Euclidean distance, the total path length can be calculated using the following formula: .

[0079] Step 2: construct the geographic distance map A geo , Load similarity graph A dtw And historical time series information is input into a pre-built load forecasting model to train the pre-built load forecasting model.

[0080] Step 2.1: Use the dilated temporal convolutional layer to process time series data. The temporal convolutional module consists of two dilated temporal convolutional layers. The first dilated temporal convolutional layer is followed by a tangent hyperbolic ( ) activation function as a filter, and another one with The activation function acts as a gate controller. The calculation process is as follows:

[0081] ,

[0082] The dilated temporal convolution layer is implemented with one-dimensional convolution to process time series data. In addition, in order to increase the receptive field to handle longer time series and reduce model complexity, the dilated convolution operation is used, and its receptive field also increases exponentially with the increase in the number of hidden layers. In order to combine spatial information later, the temporal information after time expansion is distributed to the graph convolution module and the dynamic learning graph generator to combine spatial information.

[0083] Step 2.2: The graph adjacency matrix is ​​adjusted by using learnable parameters to capture the hidden relationships between time series data. In order to make the learning graph fully represent the relationship between sites, the time series is first cut, such as Figure 2 The seasonality and trend characteristics in the input are extracted, and the time series segments after sliding average are added and then passed through a DynGraph layer (i.e. Figure 1 The DGG module in is actually a variant of a recurrent neural network to capture long-term dependencies. Figure 3 As shown in Figure 2, the learning process of the DynGraph layer is defined by the following formula:

[0084] ,

[0085] Among them, X t Indicates the trend characteristics of the sequence, X s represents the seasonal characteristics of the sequence, X w represents the weather characteristics of the sequence, AvgPool represents the sliding average operation, and α represents the sum of the trend sequence, seasonal sequence, and weather sequence after the linear layer.

[0086] After extracting information from the time series, two linear layers are used to encode the output sequence of the DynGraph layer into the latent space to obtain the embedded representation of the node. In addition, after one of the linear layers, Function as a mask to control the output information ratio and add nonlinear features. To fit more complex models:

[0087] ,

[0088] in, and Represent the value of the learned graph structure and the value of the mask at the i-th row and j-th column, respectively. represents the sigmoid activation function, Represents the dynamic learning graph matrix of the mth time period.

[0089] Step 2.3: Use the propagation layer to implement spatial graph convolution. The input of the propagation layer includes the initial time series information and a series of adjacency matrices generated by the dynamic learning graph generator. The propagation layer includes an information propagation step and an information selection step. The information propagation step is implemented by controlling hyperparameters and adjacency matrices, and the information selection step is adaptively selected for the propagation results through weighted aggregation. The final output is combined with the input of the current spatiotemporal layer as the input information of the next spatiotemporal layer.

[0090] The graph convolution module used here is a mixhop propagation layer. Given the adjacency matrix , the mixhop propagation layer includes two steps: information propagation step and information selection step, where information propagation has k steps, and each step is calculated as follows:

[0091] ,

[0092] in, is a hyperparameter used to control the retention rate of the original state information. represents the state of the previous hidden layer, Represents the adjacency matrix generated by the dynamic learning module.

[0093] The information selection step is a weighted aggregation of the k-step propagation results so that the model can make adaptive selections. The calculation process of the information selection step is as follows:

[0094] ,

[0095] in, represents the hidden state of the current layer output, represents the weight matrix, .

[0096] Step 3: Load forecasting

[0097] The time series composed of one or some sites at the time to be predicted is output to the trained load forecasting model, and the time series information of the input model is processed by the time convolution module to obtain the time series information after time expansion. The time series after time expansion processed by the time convolution module will be transmitted to the dynamic learning graph generator DGG and the spatial graph convolution module respectively; in DGG, the time series data is cut into multiple parts to generate an adjacency matrix to represent the stage-by-stage graph structure; next, the spatial graph convolution module convolves the graph structure information with the time series information processed by the time convolution module, and its output result is added to the input of the current spatiotemporal layer to obtain the input sequence of the next spatiotemporal layer.

[0098] In addition, in order to discover the sequence trends at different stages, filters of different sizes are used to convolve the time series information processed by the temporal convolution module in the L-layer spatiotemporal convolution layer. The sizes of these convolution kernels are 1×2, 1×3, 1×6 and 1×7. After being processed by the activation function and the maximum pooling layer, the feature dimensions are connected across channels to obtain the prediction model output. The temporal convolution layer of each spatiotemporal graph is as follows: Figure 1 As shown, the prediction output module projects the hidden features to the required output dimension through the fully connected layer, that is, the length of the prediction time window T′. The output of the prediction model is:

[0099] ,

[0100] Among them, z' represents the time series of the load forecasting model output, z represents the time series of the input maximum pooling layer, and f 1×k represents convolution kernels of different sizes, * represents convolution operation, max (·) represents maximum pooling operation, and concat (·) represents concatenation operation.

[0101] Example 3

[0102] This embodiment provides a subway power supply system load prediction device based on a spatiotemporal graph neural network, comprising:

[0103] Data acquisition module: used to obtain the spatial relationship between subway networks and historical load data;

[0104] A load forecasting model training module is used to train a pre-built load forecasting model through the spatial relationship and the historical load data to obtain a trained load forecasting model, wherein the pre-built load forecasting model dynamically learns according to the historical load data and continuously updates the learnable parameters in the load forecasting model;

[0105] Load forecasting result acquisition module: used to input the time series information to be predicted into the trained load forecasting model to obtain the load forecasting result, wherein the time series information to be predicted includes time information and site information.

[0106] Example 4

[0107] This embodiment provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the method for predicting the load of a subway power supply system based on a spatiotemporal graph neural network as described in Example 2 is implemented.

[0108] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present application may adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.

[0109] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0110] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0111] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0112] The embodiments of the present invention are described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the enlightenment of the present invention, ordinary technicians in this field can also make many forms without departing from the scope of protection of the purpose of the present invention and the claims, which all fall within the protection of the present invention.

Claims

1. A method for load prediction of subway power supply system based on spatiotemporal graph neural network, characterized in that: include: Obtain spatial relationships and historical load data between subway networks; The pre-built load forecasting model is trained by using the spatial relationship and the historical load data to obtain a trained load forecasting model, wherein the pre-built load forecasting model dynamically learns according to the historical load data while continuously updating the learnable parameters in the load forecasting model; The time series information to be predicted is input into a trained load prediction model to obtain a load prediction result, wherein the time series information to be predicted includes time information and site information.

2. The subway power supply system load prediction method based on spatiotemporal graph neural network according to claim 1 is characterized in that: The pre-built load forecasting model includes multiple spatiotemporal layers with residual connections and a prediction output module connected to the multiple spatiotemporal layers, wherein the output of each spatiotemporal layer is connected through a fully connected layer, wherein each spatiotemporal layer includes a temporal convolution module, a dynamic learning graph generation module and a spatial graph convolution module; the temporal convolution module includes two expanded temporal convolution layers, wherein one temporal convolution layer is connected to a tanh function and the other temporal convolution layer is connected to a sigmoid function; the dynamic learning graph generation module includes a DynGraph dynamic layer layer, and the spatial graph convolution module includes a mixhop graph network propagation layer.

3. The subway power supply system load prediction method based on spatiotemporal graph neural network according to claim 2 is characterized in that: The training of the pre-built load forecasting model according to the spatial relationship and the historical load data includes: Construct a geographic distance map based on the geographic relationships between subway networks; Construct a load similarity graph based on historical load data; The geographic distance map, the load similarity map and the historical time series information are input into a pre-built load forecasting model to obtain a trained load forecasting model.

4. The subway power supply system load prediction method based on spatiotemporal graph neural network according to claim 3 is characterized in that: The step of constructing a geographical distance map according to the geographical relationship between subway networks includes: The Euclidean distance between two sites is calculated based on the latitude and longitude of each site to obtain the geographic distance matrix D geo , define the following weighted formula to obtain the geographic distance map A geo : , in, represents the distance graph between sites i and j, represents the distance matrix between sites i and j, represents the standard deviation of the distance, Indicates the set threshold.

5. The subway power supply system load prediction method based on spatiotemporal graph neural network according to claim 3 is characterized in that: The step of constructing a load similarity graph based on historical load data includes: Arrange the time series X and Y into an n×m grid, where each point (i, j) in the grid is X i and Y j Alignment between, value is X i and Y j The absolute value of the distance difference between them, (i k , j k ) is calculated using a dynamic programming algorithm: , Where d is the Euclidean distance and the total path length is calculated using the following formula: .

6. The subway power supply system load prediction method based on spatiotemporal graph neural network according to claim 2 is characterized in that: The step of inputting the time series information to be predicted into the trained load forecasting model to obtain the load forecasting result includes: The time convolution module in the current space-time layer performs convolution processing on the space-time graph output by the previous space-time layer and the initial time series of the input model to obtain the time series after time expansion; the dynamic learning graph generation module in the current space-time layer processes the time series after time expansion to obtain the phased graph structure information; the spatial graph convolution module in the current space-time layer performs convolution processing on the graph structure information and the time series after time expansion to obtain the space-time graph output by the current space-time layer; the space-time graph output by the current space-time layer and the initial time series of the input model are added to obtain the input data of the next space-time layer, wherein the input data of the first space-time layer is the initial time series of the input model; According to the above processing steps, the time-space diagrams output by all time-space layers in the load forecasting model are obtained in sequence; The prediction output module connects all spatiotemporal graphs through a fully connected layer to obtain the predicted output sequence.

7. The subway power supply system load prediction method based on spatiotemporal graph neural network according to claim 6 is characterized in that: The convolution process of the spatiotemporal graph output by the previous spatiotemporal layer and the time series of the input model by the temporal convolution module in the current spatiotemporal layer includes: The tanh function connected after the first temporal convolution layer in the temporal convolution module is used as a filter, and the sigmoid function connected after the second temporal convolution layer is used as a gate controller. The output of the temporal convolution module is: , Among them, T1 represents the time series after convolution output by the first time convolution layer, T2 represents the time series after convolution output by the second time convolution layer, and T out Represents the convolved time series output by the temporal convolution module.

8. The subway power supply system load forecasting method based on spatiotemporal graph neural network according to claim 7 is characterized in that: The dynamic learning graph generation module in the current spatiotemporal layer processes the time series after time expansion to obtain phased graph structure information, including: Cut the historical time series and extract features from the cut historical time series; Perform sliding average on the time series after time expansion, and add the time series after sliding average to the extracted features; The summed time series is dynamically learned through the DynGraph dynamic layer. The learning process is as follows: , Among them, X t Indicates the trend characteristics of the sequence, X s represents the seasonal characteristics of the sequence, X w represents the weather characteristics of the sequence, AvgPool represents the sliding average operation, and α represents the sum of the trend sequence, seasonal sequence, and weather sequence after the linear layer; Two linear layers are used to encode the time series output by the DynGraph dynamic layer into the hidden layer to generate an adjacency matrix, where one of the two linear layers is connected to a sigmoid function as a mask. The outputs of the two linear layers and the dynamic learning module are as follows: , in, and Represent the value of the learned graph structure and the value of the mask at the i-th row and j-th column, respectively. represents the sigmoid activation function, Represents the dynamic learning graph matrix of the mth time period.

9. The subway power supply system load forecasting method based on spatiotemporal graph neural network according to claim 8 is characterized in that: The spatial graph convolution module in the current spatiotemporal layer performs convolution processing on the graph structure information and the time series after time expansion to obtain the spatiotemporal graph output by the current spatiotemporal layer, including: The adjacency matrix output by the dynamic learning module is propagated through the mixhop graph network propagation layer, wherein the information propagation has k steps, and the calculation process of each step is as follows: , in, is a hyperparameter used to control the retention rate of the original state information. represents the state of the previous hidden layer, Represents the adjacency matrix generated by the dynamic learning module; Information selection is performed on the k-step information propagation results through the mixhop graph network propagation layer. The calculation process of the information selection is as follows: , in, represents the hidden state of the current layer output, represents the weight matrix, .

10. The method for predicting load of a subway power supply system based on a spatiotemporal graph neural network according to any one of claims 6 to 9, characterized in that: The prediction output module connects all spatiotemporal graphs through a fully connected layer to obtain a predicted output sequence, including: The time series data processed by the time convolution module is convolved through four filters of different sizes, where the sizes of the convolution kernels of each filter are 1×2, 1×3, 1×6 and 1×7 respectively; The output of each filter is connected across channels after the maximum pooling operation, and the model output is obtained as follows: , Among them, z' represents the time series of the load forecasting model output, z represents the time series of the input maximum pooling layer, and f 1×k represents convolution kernels of different sizes, * represents convolution operation, max (·) represents maximum pooling operation, and concat (·) represents concatenation operation.

11. A subway power supply system load prediction device based on spatiotemporal graph neural network, characterized in that: include: Data acquisition module: used to obtain the spatial relationship between subway networks and historical load data; A load forecasting model training module is used to train a pre-built load forecasting model through the spatial relationship and the historical load data to obtain a trained load forecasting model, wherein the pre-built load forecasting model dynamically learns according to the historical load data and continuously updates the learnable parameters in the load forecasting model; Load forecasting result acquisition module: used to input the time series information to be predicted into the trained load forecasting model to obtain the load forecasting result, wherein the time series information to be predicted includes time information and site information.

Citation Information

Cited By

  • Physical-data dual-drive subway station building load prediction method

    CN119577931A

  • Physical-data dual-driven subway station building load prediction method

    CN119577931B