Short-term load forecasting method and system based on convolutional network and storage medium

Abstract

The application discloses a short-term load prediction method and system based on a convolution network, and a storage medium, and the system comprises a prediction module, which is used for inputting to-be-tested data into a trained prediction network model and obtaining predicted short-term load demand.The prediction network model comprises a feature coding module, a space-time feature representation learning module and a prediction unit; the feature coding module comprises a position coding layer, a value coding layer and a time coding layer which are arranged in parallel, and a feature adding layer.The space-time feature representation learning module is used for realizing feature learning based on different convolution kernels, and each view comprises a one-dimensional convolution layer, a causal convolution layer, a residual and regularization layer, a one-dimensional deconvolution layer and a residual and regularization layer; and the prediction unit comprises a Merge layer, a channel attention module, a channel pooling layer and a prediction layer which are sequentially connected from front to back.The application can effectively mine periodic patterns and feature rules in load time series data, has good prediction performance and high running efficiency, and has good practicability.

Description

Technical Field

[0001] This invention belongs to the technical field of short-term load forecasting, specifically relating to a short-term load forecasting method, system, and storage medium based on convolutional networks. Background Technology

[0002] Short-term load forecasting provides fundamental data for basic generation planning, system security analysis, and day-ahead market quotations, and is an important means to improve the reliability and economy of power system operation and enhance power quality on the user side. Especially under extreme heat conditions, accurately forecasting electricity load for the next one to two weeks can provide valuable reference information for urban power dispatching, preventing large-scale power outages. However, load demand is influenced by many factors, such as weather, temperature, and policies; therefore, the collected electricity load sequence data is a highly volatile and non-periodic time series. This volatility and non-periodicity make accurate forecasting challenging.

[0003] Currently, the mainstream load forecasting methods are based on deep learning technology and can be summarized into two main categories:

[0004] (1) Recurrent Neural Network-Based Methods. Commonly used methods include GRU, LSTM, and their improvements. For example, an attention module can be combined with a GRU network, or a CNN can be combined with an LSTM network to learn temporal features and make predictions. In recent years, Graph Neural Networks (GNNs) have been widely used for feature learning and modeling of data such as social networks, traffic networks, and molecular structures due to their powerful spatial network modeling capabilities. Therefore, many methods also combine GNNs with recurrent neural networks.

[0005] (2) Methods based on the Transformer framework. The Transformer framework is mainly based on the sub-attention mechanism and has a strong ability to learn data features. In recent years, many variant networks of Transformer have been proposed that can be used for load prediction. Representative models include Informer, Autoformer, and FEDformer.

[0006] In summary, although recurrent neural networks (RNNs) and Transformer-type methods demonstrate good performance in load forecasting, RNNs lag behind Transformer-type methods in terms of predictive performance and operational efficiency due to their large parameter count and slow execution speed. Therefore, this invention innovatively proposes a short-term load forecasting scheme based on convolutional networks. On the one hand, convolutional operations are more computationally efficient than Transformer models; on the other hand, convolutional models possess powerful feature mining capabilities for image data. This invention offers both good predictive performance and high operational efficiency, providing strong technical support for large-scale power system operation and scheduling. Summary of the Invention

[0007] The purpose of this invention is to provide a short-term load forecasting method, system, and storage medium based on convolutional networks, in order to solve the aforementioned problems.

[0008] This invention is mainly achieved through the following technical solutions:

[0009] The short-term load forecasting system based on convolutional networks includes a data processing module, a training module, and a prediction module. The data processing module is used to acquire historical load data of N regions at time T and form a training set. The training module is used to train a prediction network model using the training set. The prediction module is used to input the data to be tested into the trained prediction network model and obtain the predicted short-term load demand.

[0010] The prediction network model includes a feature encoding module, a spatiotemporal feature representation learning module, and a prediction unit connected sequentially from front to back. The feature encoding module includes a position encoding layer, a value encoding layer, a time encoding layer, and a feature addition layer, which are connected to the feature addition layer. The feature encoding module is used to encode the input data to add position and time information.

[0011] The spatiotemporal feature representation learning module is used to learn features based on different convolutional kernels. The spatiotemporal feature representation learning module is a multi-view mode, and different views are used to complete spatiotemporal feature learning for different cycle modes. Each view includes a one-dimensional convolutional layer, a causal convolutional layer, a residual and regularization layer, a one-dimensional deconvolutional layer, and a residual and regularization layer connected sequentially from front to back. The output of the feature addition layer is connected to the one-dimensional convolutional layer and the residual and regularization layer at the back end, respectively. The one-dimensional convolutional layer is connected to the causal convolutional layer and the residual and regularization layer at the front end, respectively.

[0012] The prediction unit includes a Merge layer, a channel attention module, a channel pooling layer, and a prediction layer connected sequentially from front to back. The Merge layer is used to fuse global spatiotemporal features in different views. The channel attention module is used to filter spatiotemporal feature information from multiple views through channel attention, providing effective spatiotemporal feature information for the prediction layer.

[0013] To better implement the present invention, the channel attention module further includes parallel max pooling layers and average pooling layers, as well as a multilayer perceptron and a Merge layer, wherein the max pooling layer and average pooling layer are respectively connected to the Merge layer through the multilayer perceptron.

[0014] To better realize the present invention, the spatiotemporal feature representation learning module is further provided with three views, and the convolution kernels are 24, 12 and 4 respectively.

[0015] To better realize the present invention, an activation layer and a Dropout operation layer are further designed between the causal convolutional layer and the residual and regularization layer, wherein the activation function is Tanh and the value of Dropout is 0.1.

[0016] This invention is mainly achieved through the following technical solutions:

[0017] The short-term load forecasting method based on convolutional networks, using the above-mentioned system, includes the following steps:

[0018] Step S1: Divide the city into N regions. The electricity load data of each region is one-dimensional time-series data. The electricity load data of the N regions are combined to form multi-dimensional time-series data, which is used as load time-series data. Obtain historical load time-series data and form a training set.

[0019] Step S2: Train the prediction network model using the training set:

[0020] Step S21: Encode the training set data using features and add location and time information;

[0021] Step S22: Set up multiple views through the spatiotemporal feature representation learning module. Each view uses a convolution kernel of a different size to realize periodic feature learning based on different convolution kernels. Different views complete spatiotemporal feature learning of different periodic patterns.

[0022] Step S23: Fuse the global spatiotemporal features from different views, and then filter the spatiotemporal feature information of the multi-view based on channel attention to provide effective spatiotemporal feature information for prediction, and finally output the predicted future load demand.

[0023] Step S3: Input the data to be tested into the trained prediction network model to obtain the predicted short-term load demand.

[0024] To better implement the present invention, the feature encoding in step S21 is further defined as follows:

[0025] The input data is encoded using the following formula:

[0026] B emb =V e +P e +T e (1)

[0027] Among them, B emb ∈R (H+T)×d ,

[0028] H is the length of the observation sequence;

[0029] T is the length of the predicted sequence;

[0030] d represents the dimension of the encoded feature;

[0031] V e It is the representation vector of the input data after padding with zeros;

[0032] P e It is the encoding of the location information of the input data;

[0033] T e The time information X of the input data time The encoding, time information X time Includes month, day, weekday, hour, and minute;

[0034] The V e P e and T e The calculation formula is as follows:

[0035] V e =Conv1d([X||X0]) (2)

[0036]

[0037] T e =MLP(X time (4)

[0038] Where: Conv1d is a one-dimensional convolutional network;

[0039] X0 is a zero-padding vector;

[0040] i is an integer value between 0 and d / 2-1;

[0041] d is the dimension of the position vector;

[0042] t is the actual position of the input value in the sequence;

[0043] MLP stands for Time-Coded Network.

[0044] To better implement the present invention, further, in step S22, the periodic feature learning in each view includes the following steps:

[0045] Step A1: Local periodic feature learning: Learning the encoded features B emb Extracting spatiotemporal features under a fixed periodicity pattern, the periodicity pattern is defined according to the size of the convolutional kernel:

[0046]

[0047] Where k is the kernel size, and stride = k in the Conv1d convolutional network, that is, feature extraction is performed by convolution in each cycle;

[0048] The dimension is

[0049] Step A2: Global Periodic Feature Learning

[0050] Mining spatiotemporal features across the entire input timescale:

[0051]

[0052] To ensure and Having the same dimensions, it is necessary to... Perform a padding operation;

[0053] Information is aggregated from different local periodic features to complete the mining of periodic features across the entire input timescale:

[0054]

[0055] The dimension of the feature vector is expanded by a one-dimensional deconvolution module to obtain the final periodic features.

[0056]

[0057]

[0058]

[0059] in:

[0060] This represents a periodic feature over the entire input timescale.

[0061] These are the regularized feature vectors;

[0062] The feature vector after deconvolution;

[0063] This is the feature vector after the dropout operation.

[0064] To better realize the present invention, step S23 further includes the following steps:

[0065] Step B1: First, use the Merge operation to fuse global spatiotemporal features from different views:

[0066]

[0067] Step B2: Calculate the channel attention weights M based on the channel attention module. c First, regarding Z global Perform max pooling and average pooling operations respectively to obtain the representation vector. and The vectors are then fused through a multilayer perceptron.

[0068]

[0069]

[0070]

[0071] Step B3: Then, for Z global Perform channel pooling:

[0072] Z c =M c ⊙Z global (16)

[0073] Z c =concat(max(Z) c ), avg(Z c (17)

[0074] Finally, prediction is achieved through a 2D convolutional layer (Conv2d) and a multilayer perceptron:

[0075] Y = MLP(Conv2d(z) c ')) (18)

[0076] in:

[0077] Z c This is the feature vector after channel pooling;

[0078] Z c′ is the concatenated vector of the maximum and mean values ​​of the channel pooling feature vectors;

[0079] Conv2d is a two-dimensional convolutional layer;

[0080] After channel pooling Z c The dimension is R (H+T)×2 .

[0081] A computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the above-described method.

[0082] The beneficial effects of this invention are as follows:

[0083] This invention designs a load data spatiotemporal feature representation learning module based on convolutional networks and introduces a multi-view mode, where different views learn spatiotemporal features of different periodic patterns. This invention also designs a prediction unit based on a channel attention mechanism suitable for multi-dimensional time-series data prediction. Channel attention filters spatiotemporal feature information from multiple views, providing effective spatiotemporal feature information for the prediction unit. This invention innovatively constructs a short-term load prediction method based on convolutional networks. This invention selects the more efficient convolutional operation, thus exhibiting high model execution efficiency. The method proposed in this invention can effectively mine periodic patterns and feature regularities in load time-series data, achieving accurate load prediction, thereby providing strong technical support for large-scale power system operation and scheduling. Attached Figure Description

[0084] Figure 1 This is a schematic diagram of the short-term load forecasting system based on convolutional networks according to the present invention.

[0085] Figure 2 This is a schematic diagram of the channel attention module based on convolutional networks in this invention. Detailed Implementation

[0086] Example 1:

[0087] Short-term load forecasting systems based on convolutional networks, such as Figure 1 As shown, it includes a data processing module, a training module, and a prediction module; the data processing module is used to acquire historical load data of N regions at time T in the past and form a training set; the training module is used to train a prediction network model using the training set; and the prediction module is used to input the data to be tested into the trained prediction network model and obtain the predicted short-term load demand.

[0088] The prediction network model includes a feature encoding module, a spatiotemporal feature representation learning module, and a prediction unit connected sequentially from front to back. The feature encoding module includes a position encoding layer, a value encoding layer, a time encoding layer, and a feature addition layer, which are connected to the feature addition layer. The feature encoding module is used to encode the input data to add position and time information.

[0089] The spatiotemporal feature representation learning module is used to learn features based on different convolutional kernels. The spatiotemporal feature representation learning module is a multi-view mode, and different views are used to complete spatiotemporal feature learning for different cycle modes. Each view includes a one-dimensional convolutional layer, a causal convolutional layer, a residual and regularization layer, a one-dimensional deconvolutional layer, and a residual and regularization layer connected sequentially from front to back. The output of the feature addition layer is connected to the one-dimensional convolutional layer and the residual and regularization layer at the back end, respectively. The one-dimensional convolutional layer is connected to the causal convolutional layer and the residual and regularization layer at the front end, respectively.

[0090] The prediction unit includes a Merge layer, a channel attention module, a channel pooling layer, and a prediction layer connected sequentially from front to back. The Merge layer is used to fuse global spatiotemporal features in different views. The channel attention module is used to filter spatiotemporal feature information from multiple views through channel attention, providing effective spatiotemporal feature information for the prediction layer.

[0091] Preferably, the channel attention module includes parallel max pooling layers and average pooling layers, as well as a multilayer perceptron and a merge layer, wherein the max pooling layers and average pooling layers are respectively connected to the merge layer through the multilayer perceptron.

[0092] Preferably, the spatiotemporal feature representation learning module has three views, and the convolution kernels are 24, 12, and 4, respectively.

[0093] Preferably, an activation layer and a Dropout operation layer are designed between the causal convolutional layer and the residual and regularization layer, wherein the activation function is Tanh and the Dropout value is 0.1.

[0094] This invention designs a load data spatiotemporal feature representation learning module based on convolutional networks and introduces a multi-view mode, where different views learn spatiotemporal features of different periodic patterns. This invention also designs a prediction unit based on a channel attention mechanism suitable for multi-dimensional time-series data prediction. Channel attention filters spatiotemporal feature information from multiple views, providing effective spatiotemporal feature information for the prediction unit. This invention innovatively constructs a short-term load prediction method based on convolutional networks. This invention selects the more efficient convolutional operation, thus exhibiting high model execution efficiency. The method proposed in this invention can effectively mine periodic patterns and feature regularities in load time-series data, achieving accurate load prediction, thereby providing strong technical support for large-scale power system operation and scheduling.

[0095] Example 2:

[0096] Short-term load forecasting methods based on convolutional networks, such as Figure 1 As shown, it includes the following steps:

[0097] 1) Data preparation: Provide historical load sequence data for the model;

[0098] 2) Feature learning: Inputting data into the load forecasting method of this invention to predict the electricity load for a future period; this process requires the following two steps:

[0099] ① Feature encoding: Encodes the input data, mainly by adding location and time information;

[0100] ② Periodic feature learning, which achieves feature learning based on different convolutional kernels;

[0101] 3) Predictive reasoning: Construct an attention module to achieve feature fusion and infer future load demand.

[0102] This invention designs a load data spatiotemporal feature representation learning module based on convolutional networks and introduces a multi-view mode, where different views complete the spatiotemporal feature learning for different periodic patterns. This invention also designs a prediction unit based on a channel attention mechanism suitable for multidimensional time-series data prediction. Channel attention filters the spatiotemporal feature information from multiple views, providing effective spatiotemporal feature information for the prediction unit.

[0103] Preferably, the present invention includes the following steps:

[0104] (1) Data preparation

[0105] The load forecasting method designed in this invention requires feature learning based on historical load data from time T in the past, and then predicts the load for the next H times. If the forecasting task requires predicting the load of N regions, then the input to the forecasting method is X = {X1, X2, ..., X...} N}.in, Similarly The lowercase letter 'x' represents the electricity load of each region at a certain time. For the forecasting method, the input load data X∈R T×N This represents T×N multidimensional time-series data. In this invention, H is set to 96. If the data sampling interval is 1 hour, then H represents 4 days.

[0106] (2) Feature learning

[0107] The data is input into the load forecasting method of this invention to predict the electricity load for a future period of time; this process requires the following two steps, which are described in detail below.

[0108] Step 1: Feature Encoding

[0109] The input data X is encoded using the following formula:

[0110] B emb =V e +P e +T e (1)

[0111] Among them, B emb ∈R (H+T)×d V e It is the representation vector of the input data after padding with zeros; P e It is the encoding of the location information of the input data; T e The time information X of the input data time ∈R (I+O)×5 The encoding, with time information mainly composed of month, day, weekday, hour, and minute, etc. e P e and T e The calculation formula is as follows:

[0112] V e =Conv1d([X||X0]) (2)

[0113]

[0114] T e =MLP(X time (4)

[0115] In formula (2), Conv1d is a one-dimensional convolutional network with N input channels and d output channels; in formula (3), t corresponds to the actual position of the input value in the sequence (i.e., which input); in formula (4), MLP is a temporal coding network consisting of 5 fully connected layers, each encoding the month, day, week, hour, and minute respectively. Then the feature vectors are summed, and the input dimension of each fully connected layer is 1, and the output dimension is d.

[0116] Step 2: Periodic Feature Learning

[0117] The periodic feature learning consists of multiple views, each corresponding to a different spatiotemporal feature module. This design takes into account the different periodicities of the load data. Each view uses a convolutional kernel of a different size, the kernel size corresponding to the period. In this invention, three views are used, with convolutional kernels of 24, 12, and 4 for each view. The execution steps of periodic feature learning in each view are described in detail below.

[0118] 1) Local periodic feature learning

[0119] The encoded feature X emb Extracting spatiotemporal features under a fixed periodicity pattern. The periodicity pattern is defined according to the size of the convolutional kernel, as follows:

[0120]

[0121] Where k corresponds to the kernel size, and the stride in the Conv1d convolutional network is k, meaning that convolution is used for feature extraction in each cycle. The dimension is

[0122] 2) Global periodic feature learning

[0123] To further explore the spatiotemporal features across the entire input timescale, which only includes local periodic features, a causal convolution module was designed:

[0124]

[0125] To ensure and Having the same dimensions, it is necessary to... Padding is performed. After the causal convolution operation, an activation layer and Dropout operation are designed. The activation function is Tanh, and the Dropout value is 0.1, i.e.:

[0126]

[0127] Its function is to aggregate information from different local periodic features in order to complete the mining of periodic features across the entire input timescale.

[0128] This invention believes that the obtained Only feature compression was implemented. Inspired by encoding and decoding frameworks, this invention aims to achieve prediction through decoding. Therefore, a one-dimensional deconvolution module was designed to expand the dimension of the feature vector, obtaining the final periodic features. The execution steps are as follows:

[0129]

[0130]

[0131]

[0132]

[0133] Residual connections are introduced before performing deconvolution, and The information is fused as shown in Equation (8). After the one-dimensional deconvolution operation, an activation layer and a Dropout operation are designed. The activation function is Tanh, and the Dropout value is 0.1. Finally, the residual connections are used to fuse the information. and B emb Information fusion yields global spatiotemporal features

[0134] During feature learning, feature learning was completed for data under different periodic patterns. Therefore, before prediction, this invention first achieves the fusion of global spatiotemporal features from different views through a Merge operation. The Merge operation for fusing global spatiotemporal features from multiple views is implemented as follows:

[0135]

[0136] (3) Predictive reasoning

[0137] Inspired by the channel attention mechanism, this invention addresses the varying importance of different dimensions of the fused periodic features for prediction. Therefore, a prediction layer is designed based on the channel attention module, and its working steps are as follows:

[0138] First, the channel attention weights M are calculated based on the channel attention module. c ,like Figure 2 As shown. First, for Z... global Perform max pooling and average pooling operations respectively to obtain the representation vector. and These two representation vectors are each processed by a multilayer perceptron, and then the output vectors are fused. The entire process can be represented as follows:

[0139]

[0140]

[0141]

[0142] Then, for Z global Channel pooling is implemented as follows:

[0143] Z c =M c ⊙Z global (16)

[0144] Z c =concat(max(Z) C ), avg(Z c (17)

[0145] In formula (16), Z after channel pooling c The dimension is R (H+T)×2 .

[0146] Finally, prediction is achieved through two-dimensional convolutional layers and a multilayer perceptron:

[0147] Y = MLP(Conv2d(Z) c ')) (18)

[0148] This invention innovatively constructs a short-term load forecasting method based on convolutional networks. The invention selects convolutional operations, which have higher execution efficiency, thus achieving high model execution efficiency. The proposed method can effectively mine periodic patterns and characteristic regularities in load time-series data, achieving accurate load forecasting, thereby providing strong technical support for large-scale power system operation and scheduling.

[0149] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Any simple modifications or equivalent changes made to the above embodiments based on the technical essence of the present invention shall fall within the protection scope of the present invention.

Claims

1. A short-term load forecasting method based on convolutional networks, characterized in that, Includes the following steps: Step S1: Divide the city into N regions. The electricity load data of each region is one-dimensional time-series data. The electricity load data of the N regions are combined to form multi-dimensional time-series data, which is used as load time-series data. Obtain historical load time-series data and form a training set. Step S2: Train the prediction network model using the training set: Step S21: Encode the training set data using features and add location and time information; Step S22: Set up multiple views through the spatiotemporal feature representation learning module. Each view uses a convolution kernel of a different size to realize periodic feature learning based on different convolution kernels. Different views complete spatiotemporal feature learning of different periodic patterns. Step S23: Fuse the global spatiotemporal features from different views, and then filter the spatiotemporal feature information of the multi-view based on channel attention to provide effective spatiotemporal feature information for prediction, and finally output the predicted future load demand. Step S3: Input the data to be tested into the trained prediction network model to obtain the predicted short-term load demand; In step S22, the periodic feature learning in each view includes the following steps: Step A1: Local periodic feature learning: Learning the encoded features B emb Extracting spatiotemporal features under a fixed periodicity pattern, the periodicity pattern is defined according to the size of the convolutional kernel: （5） in, k The kernel size is [size]. Conv 1 d Stride size in convolutional networks stride=k That is, convolution is used for feature extraction in each cycle; The dimension is ; Step A2: Global Periodic Feature Learning Mining spatiotemporal features across the entire input timescale: （6） To ensure and Having the same dimensions, it is necessary to... Perform a padding operation; Information is aggregated from different local periodic features to complete the mining of periodic features across the entire input timescale: （7） The dimension of the feature vector is expanded by a one-dimensional deconvolution module to obtain the final periodic features. : （8） （9） （10） （11） in: This represents a periodic feature over the entire input timescale. These are the regularized feature vectors; The feature vector after deconvolution; The feature vector after the dropout operation; Step S23 includes the following steps: Step B1: First, use the Merge operation to fuse global spatiotemporal features from different views: （12） Step B2: Calculate channel attention weights based on the channel attention module M c First of all, Perform max pooling and average pooling operations respectively to obtain the representation vector. and The vectors are then processed through a multilayer perceptron and fused to output the following: （13） （14） （15） Step B3: Then, for Perform channel pooling: （16） （17） Finally, after two-dimensional convolutional layers Conv2d And multilayer perceptron for prediction: （18） in: Z c This is the feature vector after channel pooling; Z c ' This is a concatenated vector of the maximum and mean values ​​of the channel pooling feature vectors. Conv2d It is a two-dimensional convolutional layer; After channel pooling Z c The dimension is .

2. The short-term load forecasting method based on convolutional networks according to claim 1, characterized in that, The feature encoding in step S21 is as follows: The input data is encoded using the following formula: （1） in, , H The length of the observation sequence; T To predict sequence length; d For encoding feature dimensions; V e It is the representation vector of the input data after padding with zeros; P e It is the encoding of the location information of the input data; T e The time information X of the input data time The encoding, time information X time Includes month, day, weekday, hour, and minute; The V e , P e and T e The calculation formula is as follows: （2） （3） （4） in: Conv 1 d It is a one-dimensional convolutional network; X 0 is a vector with zeros as its complement; i It is an integer value between 0 and d / 2-1; d The dimension of the position vector; t It is the actual position of the input value in the sequence; MLP It is a time-coded network.

3. A short-term load forecasting system based on convolutional networks, used to execute the short-term load forecasting method based on convolutional networks as described in claim 1; characterized in that, It includes a data processing module, a training module, and a prediction module; the data processing module is used to acquire historical load data of N regions at time T in the past and form a training set; the training module is used to train a prediction network model using the training set; and the prediction module is used to input the data to be tested into the trained prediction network model and obtain the predicted short-term load demand. The prediction network model includes a feature encoding module, a spatiotemporal feature representation learning module, and a prediction unit connected sequentially from front to back. The feature encoding module includes a position encoding layer, a value encoding layer, a time encoding layer, and a feature addition layer, which are connected to the feature addition layer. The feature encoding module is used to encode the input data to add position and time information. The spatiotemporal feature representation learning module is used to learn features based on different convolutional kernels. The spatiotemporal feature representation learning module is a multi-view mode, and different views are used to complete spatiotemporal feature learning for different cycle modes. Each view includes a one-dimensional convolutional layer, a causal convolutional layer, a residual and regularization layer, a one-dimensional deconvolutional layer, and a residual and regularization layer connected sequentially from front to back. The output of the feature addition layer is connected to the one-dimensional convolutional layer and the residual and regularization layer at the back end, respectively. The one-dimensional convolutional layer is connected to the causal convolutional layer and the residual and regularization layer at the front end, respectively. The prediction unit includes a Merge layer, a channel attention module, a channel pooling layer, and a prediction layer connected sequentially from front to back. The Merge layer is used to fuse global spatiotemporal features in different views. The channel attention module is used to filter spatiotemporal feature information from multiple views through channel attention, providing effective spatiotemporal feature information for the prediction layer.

4. The short-term load forecasting system based on convolutional networks according to claim 3, characterized in that, The channel attention module includes parallel max pooling and average pooling layers, as well as a multilayer perceptron and a Merge layer. The max pooling and average pooling layers are connected to the Merge layer through the multilayer perceptron.

5. The short-term load forecasting system based on convolutional networks according to claim 3, characterized in that, The spatiotemporal feature representation learning module has three views, with convolutional kernels of 24, 12, and 4 respectively.

6. The short-term load forecasting system based on convolutional networks according to claim 3, characterized in that, An activation layer and a Dropout operation layer are designed between the causal convolutional layer and the residual and regularization layer. The activation function is Tanh, and the Dropout value is 0.

1.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method described in any one of claims 1-2.

Images