Day-ahead electricity price prediction method based on parallel multi-dimensional attention mechanism

Abstract

Provided is a day-ahead electricity price prediction method based on a parallel multi-dimensional attention mechanism, which method belongs to the technical field of day-ahead electricity price prediction in a power market. The maximum mutual information coefficient is used to select variables having a certain relevance with an electricity price, so as to assist in predicting future electricity price data; next, a decomposition algorithm is used to decompose an original electricity price signal; and then, electricity price sub-components obtained by means of decomposition and the electricity-price-related variables are inputted into an MBI-IPMDA-PBISA deep learning model, so as to predict a future electricity price. In this way, an MBI-IPMDA-PBISA deep learning model can fully learn a complex association relationship between electricity price sub-components and electricity-price-related variables and a future electricity price, the problem of error superimposition caused by separately predicting the electricity price sub-components and then superimposing same to obtain the future electricity price is also avoided, and the problem of the accuracy of existing attention mechanisms during prediction being low is solved, thereby improving the prediction accuracy.

Description

Day-ahead electricity price prediction method based on parallel multi-dimensional attention mechanism TECHNICAL FIELD

[0001] The application belongs to the technical field of day-ahead electricity price prediction in the electricity market, and specifically relates to a day-ahead electricity price prediction method based on a parallel multi-dimensional attention mechanism. BACKGROUND

[0002] After a large amount of clean energy such as wind power and photovoltaic power is added to the electricity market, the electricity price in the electricity market fluctuates dramatically, thereby increasing the difficulty of electricity price prediction, so how to accurately predict the day-ahead electricity price in the case of a high proportion of new energy added to the electricity market has become a major problem that researchers at home and abroad are closely concerned about.

[0003] With the development of deep learning and data processing technology, in recent years, many studies have begun to combine multiple algorithms to form a complex prediction framework. There is a lot of feature information related to future electricity price data in electricity price data and electricity price related variable data, and how to mine useful deep features to predict future electricity price data is very important.

[0004] Attention mechanism is a mechanism that simulates the different attention levels of humans to important features to select more critical information and improve model accuracy. In 2017, Google proposed a self-attention mechanism (SA), which simulates the attention selectivity of biological vision to capture the key information of the input sequence itself and has strong adaptive ability, suitable for processing long-term dependencies of input sequences. In 2018, Jie Hu et al. proposed a channel attention network (SE-Net), which selectively adjusts the importance of different feature channels using a channel attention mechanism, so that the network pays more attention to useful feature channels in the feature map to improve the performance of the entire network model; In 2018, Sanghyun WOO et al. proposed a convolutional block attention mechanism module (CBAM), which uses a channel attention module (CAM) and a spatial attention module (SAM) to more comprehensively improve the attention level of important features; In 2021, Qibin Hou et al. proposed a coordinate attention mechanism (CA), which integrates feature map position information into the channel attention mechanism to improve the attention level of more important feature points in the feature map, rather than only focusing on important feature regions of the feature map.

[0005] However, since the SE attention mechanism globally processes the information in the channel, it only focuses on the importance of the channel and ignores the information interaction in the space. The CBAM focuses on the information interaction of the channel and the space through the channel attention module and the spatial attention module in series. The spatial attention model uses a 7*7 convolution to extract spatial features after pooling the channels, which can only capture local information and will cause some long-range dependencies to be ignored. The above methods cannot be considered comprehensively, resulting in low prediction accuracy of the existing attention mechanism. SUMMARY

[0006] The purpose of the present application is to solve the problem of low prediction accuracy caused by the fact that the existing attention mechanisms do not consider comprehensively during prediction. The present application proposes a day-ahead electricity price prediction method based on parallel multi-dimensional attention mechanism. The method selects variables related to electricity price by using the maximum mutual information coefficient to assist in predicting future electricity price data. Secondly, the original electricity price signal is decomposed using a decomposition algorithm. Then, the electricity price sub-components and electricity price related variables obtained by decomposition are input into the MBI-IPMDA-PBISA deep learning model to predict future electricity prices. This can make the prediction model fully learn the complex correlation between the electricity price sub-components, the electricity price related variables and the future electricity price, and avoid the error accumulation problem caused by separately predicting the electricity price sub-components and then superimposing the future electricity price. The method solves the problem of low prediction accuracy of the existing attention mechanism and improves the prediction accuracy.

[0007] Compared with the prior art, the present application comprehensively considers the different deep features contained in the input electricity price and electricity price related variables, uses different feature extraction modules to mine the feature information, and finally proposes the MBI-IPMDA-PBISA deep learning model.

[0008] To solve the above technical problems, the present application provides a day-ahead electricity price prediction method based on parallel multi-dimensional attention mechanism, which comprises:

[0009] S1, collecting electricity price data and electricity price related variables, and performing data preprocessing on the collected data to construct an electricity price prediction dataset;

[0010] S2, calculating the maximum mutual information coefficient between the electricity price data and the electricity price related variables based on the electricity price prediction dataset, and selecting the electricity price related variables related to the electricity price by using the maximum mutual information coefficient;

[0011] S3, decomposing the electricity price data using a decomposition algorithm to obtain electricity price sub-components;

[0012] S4, constructing an MBI-IPMDA-PBISA deep learning model;

[0013] S5, input and output division is performed on the electricity price sub-component and the electricity price related variables screened in S2 to obtain a dataset sample, each sample containing the electricity price sub-component and the electricity price related variables and the future electricity price true value, the electricity price sub-component and the electricity price related variables in each sample are merged respectively to obtain an electricity price sub-component input matrix and an electricity price related variable input matrix, and finally each sample contains the electricity price sub-component input matrix and the electricity price related variable input matrix and the future electricity price true value;

[0014] S6, the samples are divided into a training set and a test set, the training set is input into the MBI-IPMDA-PBISA deep learning model for training, and then the test set is input into the trained MBI-IPMDA-PBISA deep learning model for testing to predict the future electricity price.

[0015] Preferably, the data preprocessing includes processing of abnormal values in the electricity price prediction dataset and standardization processing of the electricity price prediction dataset.

[0016] Preferably, the processing of abnormal values includes: for data missing values, a sliding window filling method is used for supplement; for the same time node and repeated data values, one of the data is directly used, and for the same time node and different data values, the average value of the two data is taken for subsequent calculation; for data abnormal values, 3δ criterion or Grubbs test method is used to test the electricity price data and the electricity price related variable data, if it is determined that a point is an abnormal value data, it is deleted and then treated as a data missing value.

[0017] Preferably, the MBI-IPMDA-PBISA deep learning model includes:

[0018] Sa, for the electricity price sub-component input branch, a 2-layer Inception module is used in series to extract multi-scale spatial features of the electricity price sub-component, a plurality of multi-scale spatial feature matrices are obtained, and the plurality of multi-scale spatial feature matrices are stacked in the third dimension to obtain multi-scale spatial features of the electricity price sub-component;

[0019] Sb, a parallel multi-dimensional attention mechanism is used in series to adjust the attention weight of each dimension in the multi-scale spatial features of the electricity price sub-component;

[0020] Sc, for the electricity price related variable input branch, an Inception module and a BiLSTM module are used in parallel to simultaneously extract spatial features and time sequence features of the electricity price related variables, and the spatial features and the time sequence features are fused to obtain space-time features of the electricity price related variables;

[0021] Sd, a self-attention mechanism is used in series to adjust the internal attention weight in the space-time features of the electricity price related variables;

[0022] The multi-scale spatial features of the electricity price sub-components with adjusted attention weights in each dimension and the spatio-temporal features of the electricity price related variables with adjusted internal attention weights are fused, and then further integrated through a fully connected layer and a regression layer to obtain the output of the MBI-IPMDA-PBISA deep learning model, i.e., the predicted future electricity price.

[0023] Preferably, the Sa includes: inputting the electricity price sub-components into the first Inception module of the electricity price sub-component input branch, and the second Inception module of this branch and the Inception module in the electricity price related variable input branch have the same calculation process as the Inception module in the electricity price related variable input branch, and the multi-scale spatial feature matrix extracted by the second Inception module of the electricity price sub-component input branch is input into the series-parallel multi-dimensional attention mechanism; wherein the calculation process of the first Inception module of the electricity price sub-component input branch is as follows:

[0024] In the formula, F p is the Danish electricity price decomposition matrix, is the weight matrix of the i-th convolution kernel of the 1×1 convolution branch; are the weight matrices of the i-th convolution kernel of the first layer 3×3, 5×5, and 7×7 convolution, respectively; are the weight matrices of the i-th convolution kernel of the second layer 3×3, 5×5, and 7×7 convolution, respectively; is the bias of the i-th convolution kernel of the 1×1 convolution; are the biases of the i-th convolution kernel of the first layer 3×3, 5×5, and 7×7 convolution, respectively; are the biases of the i-th convolution kernel of the second layer 3×3, 5×5, and 7×7 convolution, respectively; is the output feature of the 1×1 convolution branch; are the output features of the first layer 3×3, 5×5, and 7×7 convolution, respectively; are the output features of the second layer 3×3, 5×5, and 7×7 convolution, respectively; and δr is the Relu activation function; the feature matrices extracted by the four convolution branches are concatenated to obtain the output multi-scale spatial feature matrix Similarly, the multi-scale spatial feature matrices output by the second Inception module in the electricity price sub-component input branch and the Inception module in the electricity price related variable input branch are and

[0025] Preferably, the Sb includes:

[0026] Sb1, using global max pooling and global average pooling to process the multi-scale spatial feature matrix. The spatial information is compressed and aggregated to obtain a feature map of size C×1×1. Two 1×1 convolutional layers are set for scaling learning, and after activation by the Sigmoid function, the attention weights in the first dimension are obtained. The calculation process is shown in the following formula:

[0027] In the formula: GMP and GAP are global max pooling and global average pooling operations, respectively; It is the dot product of vectors; and The weights of the i-th 1×1 convolutional kernel in the first and second layers along dimension C; and The bias is the bias of the i-th 1×1 convolution kernel in the first and second layers; This is the feature output after processing the dot product of global max pooling and global average pooling; and f CA The output is the result of feature learning through the first and second 1×1 convolution layers, where σ is the Sigmoid activation function.

[0028] Sb2. Two one-dimensional max pooling and one-dimensional average pooling methods are used to compress and aggregate the vertical and horizontal spatial information of the multi-scale spatial feature matrix, respectively, to obtain feature maps with outputs of C×H×1 and C×1×W. The calculation process is shown in the following formula:

[0029] In the formula: and for The output of the j-th channel in the H dimension after average pooling and max pooling; and for The output of the j-th channel in the W dimension after average pooling and max pooling; τ represents the max pooling and average pooling operations with different spatial ranges for each channel;

[0030] Two 1×1 convolutional layers are used for scaling learning, and then activated by the Sigmoid function to obtain the vertical and horizontal attention weights. The calculation process is shown in the following formula:

[0031] In the formula: and The weights of the i-th 1×1 convolutional kernel in the first and second layers along the H dimension; and is the bias of the i-th 1x1 convolution kernel in the H dimension of the first layer and the second layer; and is the weight of the i-th 1x1 convolution kernel in the W dimension of the first layer and the second layer; and is the bias of the i-th 1x1 convolution kernel in the W dimension of the first layer and the second layer; and f HA is the output learned by the 1x1 convolution of the first layer and the second layer on the H dimension feature; and f WA is the output learned by the 1x1 convolution of the first layer and the second layer on the W dimension feature;

[0032] Sb3, the attention weight obtained in different dimensions is multiplied with the original multi-scale spatial feature matrix to obtain a multi-dimensional attention feature matrix f DA , as shown in the following formula:

[0033] As a preferred, the Sc comprises:

[0034] For the electricity price related variable input branch, an Inception and BiLSTM parallel structure is adopted to simultaneously extract the spatial features and the time sequence features of the electricity price related variables; the electricity price related variable input matrix is transformed into a 1-dimensional long vector by a flattening layer and is input into a BiLSTM module, the contact between the historical data and the future data is strengthened through the forward and backward propagation rules, and the time sequence features of the electricity price related variables are extracted, and the calculation process is as shown in the following formula:

[0035] In the formula: and are the outputs of the forward hidden layer and the backward hidden layer at t moment respectively; and are the weight matrices of the input layer mapped to the forward hidden layer and the backward hidden layer respectively; and are the weight matrices of the previous moment output of the forward hidden layer and the backward hidden layer mapped to the current moment respectively; b (1) and b (2) are the biases of the forward hidden layer and the backward hidden layer respectively; W1 and W2 are the weight matrices of the output of the forward hidden layer and the backward hidden layer mapped to the output layer respectively; δt is a Tanh activation function; is the output of the electricity price related variable matrix after the flattening layer; represents the output of the output layer. Similarly, the time sequence feature f v,1 is input into the second BiLSTM network of this branch to obtain the time sequence feature f is used subsequently representative

[0036] spatial features extracted by the inception module temporal features extracted by the BiLSTM module after being flattened spatio-temporal features f of the electricity price related variable data v The calculation process is as follows:

[0037] wherein, denotes concatenation in the first dimension.

[0038] As a preferred, the Sd includes:

[0039] By self-attention mechanism, different position information in the spatio-temporal features of the electricity price related variable data is associated and interacted with each other, the dependency relationship therein is more comprehensively learned, and important information therein is focused on. The calculation process is as follows:

[0040] Randomly initialize 4 sets of weight matrices spatio-temporal features f of the electricity price related variable data v Multiply the 4 sets of weight matrices to obtain 4 sets of Q i , K i , and V i (i∈0,1,2,3); perform inner product of the matrix Q i and K i , normalize by the softmax function to obtain a correlation weight matrix, then perform weighted summation of the matrix V i with the correlation weight matrix, thereby obtaining a weight coefficient Z; finally, multiply Z and an initialized matrix W0 to obtain the spatio-temporal features f containing the self-attention mechanism. vSA .

[0041] In the formula: denotes the initialized weight matrix, Q i , K i , and V i denote query vector, key vector, and value vector respectively, d k denotes the dimension of the key vector, Z denotes the weight coefficient, and W 0 denotes the weight matrix.

[0042] As preferred, the data set samples are divided into a training set and a test set, the training set is first input into the MBI-IPMDA-PBISA deep learning model for training, a trained MBI-IPMDA-PBISA deep learning model is obtained, then the test set data is input into the trained MBI-IPMDA-PBISA deep learning model to predict future electricity prices, the prediction result is compared with the true value, and the prediction accuracy and prediction error of the test set are obtained. In engineering application, historical data is input into the trained MBI-IPMDA-PBISA deep learning model, and the future electricity price data to be predicted can be obtained.

[0043] The beneficial effects of the present application are:

[0044] 1、The parallel multi-dimensional attention mechanism is proposed to avoid the problem that the SE attention mechanism only focuses on the importance of each channel and ignores the information interaction in space, and the problem that CBAM can only capture local information by connecting the channel attention module and the spatial attention module to focus on the information interaction of channels and space. The parallel multi-dimensional attention mechanism uses the idea that "two one-dimensional global pooling operations are used to extract feature information in the vertical and horizontal directions respectively, and the two independent direction perception feature maps can capture the long-range dependency relationship in different directions of the input feature map", and combines the advantages of CBAM which uses average pooling and maximum pooling to aggregate feature information, fully considers different feature extraction modules for feature information mining, and improves the prediction accuracy of the MBI-IPMDA-PBISA model for prediction data.

[0045] 2、The present scheme considers the related variables of electricity price to assist in predicting future electricity price, and effectively eliminates the related variables with low correlation with electricity price by using the maximum household information coefficient, so as to reduce the calculation amount of the deep learning prediction model.

[0046] 3、The MBI-IPMDA-PBISA deep learning model proposed in the present scheme first extracts deep features in the electricity price and the electricity price related variables respectively, and then performs feature fusion to predict the future electricity price, so that the prediction model can fully learn the complex correlation between the electricity price sub-components and the electricity price related variables and the future electricity price, and avoid the error accumulation problem caused by separately predicting the electricity price sub-components and then superimposing the future electricity price.

[0047] 4. The scheme extracts the spatial features of the input data by using three convolution layers with different convolution window sizes of 1*1, 3*3 and 5*5 and a maximum pooling layer, the receptive fields of different convolution window sizes are different, the extracted spatial features are different, the features extracted by the pooling layer are different from those extracted by the convolution layer, and the combination of the spatial feature matrix with multiple channels can complement the feature extraction advantages of different convolution layers and pooling layers and enhance the feature extraction performance; meanwhile, the scheme adopts the idea that the spatial features extracted by different convolution window sizes in the Inception deep learning model are different, and the three convolution layers with different convolution window sizes of 1*1, 3*3 and 5*5 in the Inception basic model are used to increase the feature diversity. The maximum pooling layer will ignore some element values when extracting features, which is more likely to cause overfitting problem and will excessively improve the accuracy of the training set, thereby leading to the decline of the test accuracy. Therefore, the Inception deep learning model used in the application uses a 7*7 convolution module to replace the maximum pooling layer, and in order to expand the receptive field of the convolution and improve the network depth, the 3*3, 5*5 and 7*7 convolutions are set to two layers; furthermore, the LSTM network can only process data in one direction, often ignoring the information hidden in the future time. The BiLSTM network is composed of forward and backward LSTM networks, which can well utilize the future information while obtaining the past information, which can further improve the prediction accuracy of the data with time sequence. BRIEF DESCRIPTION OF DRAWINGS

[0048] Other features, objects and advantages of the present application will become more apparent from the following detailed description of non-limiting embodiments made with reference to the accompanying drawings. The drawings are for purposes of illustrating the preferred embodiments only and are not to be construed as limiting the application. Moreover, the same reference numerals are used throughout the drawings to denote the same components.

[0049] Fig. 1 is a schematic diagram of the parallel multi-dimensional attention mechanism of the application;

[0050] Fig. 2 is a schematic diagram of the MBI-IPMDA-PBISA deep learning model structure of the application;

[0051] Fig. 3 is a schematic diagram of the data change curve containing missing values of the application;

[0052] Fig. 4 is a diagram of the normalized electricity price and its related variable data of the application;

[0053] Fig. 5 is a diagram of the maximum mutual information coefficient value between DK1 electricity price and other related variables of the application;

[0054] Fig. 6 is a schematic diagram of the DK1 electricity price data group decomposition result of the application;

[0055] Fig. 7 is a schematic diagram of the input and output of the sliding window construction model of the present application;

[0056] Fig. 8 is a schematic diagram of the prediction results of the MBI-IPMDA-PBISA deep learning model of the present application;

[0057] Fig. 9 is a schematic diagram of the day-ahead electricity price prediction error of the prior art model with different attention mechanisms. DETAILED DESCRIPTION

[0058] In order to make the purpose, technical solutions and advantages of the present application clearer and more apparent, the present application will be further described in detail below in conjunction with the drawings and examples. It should be understood that the specific embodiments described herein are only the best mode of the present application, which are used to explain the present application and do not limit the protection scope of the present application. All other embodiments obtained by those skilled in the art without making creative efforts fall within the protection scope of the present application.

[0059] Example 1: As shown in Fig. 1, the day-ahead electricity price prediction method based on parallel multi-dimensional attention mechanism includes:

[0060] Step 1: Collect electricity price data P and electricity price related variables such as power generation, electricity consumption, electricity trading volume, etc. Vm, where m represents the number of electricity price related variables.

[0061] Step 2: Process the abnormal values in all data.

[0062] For missing data values, the sliding window filling method is used for supplementation (such as the data processing process in the implementation example)

[0063] For data repetition values, which are generally caused by recording errors, the same data value can be directly used, and for different data values, the average of the two data can be taken for subsequent calculation. (Data repetition is rare, and data repetition but different data values may be caused by special circumstances, i.e. the electricity price changes in a period of time when it should remain unchanged. The average value can be taken to make it consistent with other data formats)

[0064] For data abnormal values (sudden large distortion of a point in the data due to recording error), 3δ criterion or Grubbs test method can be used to test the electricity price data and electricity price related variable data. If it is determined that a point is an abnormal value data, it is deleted and then treated as a missing value.

[0065] Step 3: Standardize all data to unify the dimension size of electricity price and electricity price related variables.

[0066] Step 4: Calculate the maximum mutual information coefficient between the electricity price and the electricity price related variables, and select the variables related to the electricity price to assist in predicting future electricity price data.

[0067] The maximum mutual information coefficient ranges from 0 to 1, and the closer the value is to 1, the greater the correlation between the current electricity price related variable and the historical electricity price, and the closer the value is to 0, the smaller the correlation between the current electricity price related variable and the historical electricity price. How to set the threshold has a great influence on the subsequent overall experiment, and setting the threshold too high will result in too few electricity price related variables being selected, causing some important feature information contained in other electricity price related variables to be ignored, and setting the threshold too low will result in too much redundant information of the electricity price related variables being input into the deep learning model, resulting in an increase in the calculation amount and prediction difficulty of the deep learning model.

[0068] In the implementation example, the selection of the electricity price related variables is made by referring to other related research and comprehensive consideration, and finally the threshold is determined to be 0.2.

[0069] Step 5: Decompose the electricity price signal using a signal decomposition algorithm to obtain electricity sub-components IMF1-IMFn.

[0070] Step 6: Construct the MBI-IPMDA-PBISA deep learning model while performing MBI-IPMDA-PBISA deep learning model input and output division on the electricity sub-components and the electricity price related variables to obtain the electricity sub-component input matrix F p and the related variable input matrix Fv, wherein p and v respectively represent the sample number of the electricity sub-component input matrix and the related variable input matrix.

[0071] Step 7: Divide all samples into a training set and a test set.

[0072] Step 8: Input the training set data into the MBI-IPMDA-PBISA deep learning model to train the model.

[0073] As shown in FIG. 2, it can be understood that, for the electricity price sub-component input branch of the MBI-IPMDA-PBISA deep learning model, since the historical electricity price data is obtained after decomposition, a plurality of electricity price sub-components are obtained, the time sequence correlation information of the original electricity price data is changed, and there is spatial correlation information between the electricity price sub-components, therefore, a 2-layer Inception module is used in series to extract multi-scale spatial features, different sizes of convolution kernels can extract different scale features, the extracted features are stacked to obtain multi-scale electricity price spatial features with richer feature information, so as to enhance the feature extraction capability of the model, and then a PMDA module (parallel multi-dimensional attention mechanism) is connected in series to adjust the attention weight on each dimension of the multi-scale electricity price spatial features extracted by the Inception module.

[0074] For the electricity price related variable input branch, since the time sequence correlation information is contained in each related variable, and there is spatial correlation information between the related variables, an Inception module and a BiLSTM module are used in parallel to simultaneously extract the spatial features and the time sequence features of the electricity price related variables, the spatial features and the time sequence features are fused to obtain the space-time features of the electricity price related variables, and then a SA module (self-attention mechanism) is connected in series to adjust the internal attention weight of the space-time features.

[0075] The multi-scale electricity price spatial features adjusted by the PMDA module and the space-time features of the electricity price related variables adjusted by the SA module are fused, and then input to a fully connected layer for further integration, and finally output the future electricity price prediction value.

[0076] The Inception basic model was proposed by the Google team in the ImageNet competition in 2014, which uses different sizes of convolution windows and pooling to extract spatial feature matrices with different feature rules in the input picture, and then stacks them to give to the next module, which well preserves the different regularities of each spatial feature matrix and increases the width of the network, and has strong spatial feature extraction capability.

[0077] The network uses three convolution layers with different sizes of 1x1, 3x3 and 5x5 convolution windows and a maximum pooling layer to extract spatial features of the input data. The receptive fields of different convolution window sizes are different, and the extracted spatial features are different. The features extracted by the pooling layer are different from those extracted by the convolution layer. Combining them into a multi-channel spatial feature matrix can complement the feature extraction advantages of different convolution layers and pooling layers, and enhance their feature extraction performance.

[0078] The MBI-IPMDA-PBISA deep learning model adopts the different spatial feature extraction ideas of different convolution window sizes in the Inception deep learning model, and the three convolution window size different convolutions of 1*1, 3*3 and 5*5 in the Inception basic model are used to increase the feature diversity.

[0079] Since the LSTM network can only process data in one direction, it often ignores the hidden information of future time. The BiLSTM network is composed of forward and backward LSTM networks, which can make good use of future information while obtaining past information. This can further improve the prediction accuracy of time series data.

[0080] The BiLSTM network is composed of an input layer, an output layer, a forward hidden layer and a backward hidden layer. The input data of the input layer is transmitted to the forward hidden layer and the backward hidden layer at the same time, and the data of the output layer is composed of the forward hidden layer and the backward hidden layer.

[0081] The calculation process of the MBI-IPMDA-PBISA deep learning model is as follows: for the electricity price sub-component matrix in the input 1, first use 2 Inception modules to extract multi-scale features, different size convolution kernels can extract different scale features, and stack the extracted features to obtain electricity price spatial features with richer feature information, so as to enhance the feature extraction ability of the model. The calculation process of the first Inception module in the electricity price input branch is shown in formula (1), and the calculation process of the second Inception module and the Inception module in the electricity price related variable input branch is the same.

[0082] In the formula: F p is the Danish electricity price decomposition matrix, is the weight matrix of the i-th convolution kernel of the 1*1 convolution branch; are the weight matrices of the i-th convolution kernel of the first layer 3*3, 5*5 and 7*7 convolution, respectively; are the weight matrices of the i-th convolution kernel of the second layer 3*3, 5*5 and 7*7 convolution, respectively; is the bias of the i-th convolution kernel of the 1*1 convolution; Bias of the i-th convolution kernel of 3x3, 5x5, 7x7 convolution in the first layer, respectively; Bias of the i-th convolution kernel of 3x3, 5x5, 7x7 convolution in the second layer, respectively; Output feature of 1x1 convolution branch; Output feature of 3x3, 5x5, 7x7 convolution in the first layer, respectively; Output feature of 3x3, 5x5, 7x7 convolution in the second layer, respectively; δr is a Relu activation function; the feature matrix extracted by the four convolution branches is connected by Concat to obtain the output multi-scale spatial feature matrix Similarly, the multi-scale spatial feature matrix output by the second Inception module in the electricity price decomposition matrix input branch and the Inception module in the electricity price related variable input branch is and

[0083] The multi-scale spatial feature matrix extracted by the second Inception module in the electricity price decomposition matrix input branch is input into the PMDA module, which first compresses and aggregates the spatial information of the multi-scale spatial feature matrix through global maximum pooling and global average pooling to obtain a feature map with an output size of Cx1x1, and then sets two 1x1 convolution layers for scaling learning and obtains the attention weight in the first dimension (channel) after Sigmoid function activation, and the calculation process is shown in formula (2).

[0084] In the formula: GMP and GAP are global maximum pooling and global average pooling operations; is a vector dot product; and are the weights of the i-th 1x1 convolution kernel in the first layer and the second layer in the C dimension; and are the biases of the i-th 1x1 convolution kernel in the first layer and the second layer; is the feature output after global maximum pooling and global average pooling dot processing; and f CA is the output after feature learning by the first layer and the second layer 1x1 convolution, and σ is a Sigmoid activation function. Then, two one-dimensional maximum pooling and one-dimensional average pooling are used to compress and aggregate the vertical direction spatial information and horizontal direction spatial information of the multi-scale spatial feature matrix, respectively, to obtain feature maps with output sizes of CxHx1 and Cx1xW, and the calculation process is shown in formulas (3) and (4). Then, two 1x1 convolution layers are set for scaling learning and the attention weights in the second dimension (vertical) and the third dimension (horizontal) are obtained after Sigmoid function activation, and the calculation process is shown in formula (5).

[0085] wherein: and are the output of the jth channel in the H dimension after average pooling and max pooling; and are the output of the jth channel in the W dimension after average pooling and max pooling; τ represents the max pooling and average pooling operation with different spatial ranges for each channel; and are the weights of the i th 1 × 1 convolution kernel in the H dimension of the first layer and the second layer; and are the biases of the i th 1 × 1 convolution kernel in the H dimension of the first layer and the second layer; and are the weights of the i th 1 × 1 convolution kernel in the W dimension of the first layer and the second layer; and are the biases of the i th 1 × 1 convolution kernel in the W dimension of the first layer and the second layer; and f HA are the outputs learned by the 1 × 1 convolution of the first layer and the second layer for the H dimension features; and f WA are the outputs learned by the 1 × 1 convolution of the first layer and the second layer for the W dimension features.

[0086] The attention weights obtained in different dimensions are multiplied with the original multi-scale spatial feature matrix to obtain a multi-dimensional attention feature matrix f DA as shown in equation (6).

[0087] For the electricity price related variable input branch, an Inception and BiLSTM parallel structure is adopted to simultaneously extract the spatial features and time series features of the electricity price related variables. The electricity price related variable data is transformed into a 1-dimensional long vector by a flattening layer and input into the BiLSTM module. The connection between the historical data and the future data is strengthened through the forward and backward propagation rules, and the time series features of the electricity price related variables are extracted, the calculation process is shown in equations (7) and (8).

[0088] wherein: and are the outputs of the forward hidden layer and the backward hidden layer at time t respectively; and These are the weight matrices that map the input layer to the forward hidden layer and the backward hidden layer, respectively. and These are the weight matrices mapping the outputs of the forward and backward hidden layers from the previous time step to the current time step, respectively; b (1) and b (2) W1 and W2 are the biases of the forward and backward hidden layers, respectively; W1 and W2 are the weight matrices mapping the outputs of the forward and backward hidden layers to the output layer, respectively; δt is the Tanh activation function. This is the output of the electricity price-related variable matrix after it has been flattened. This represents the output of the output layer. Similarly, the time sequence feature f is... v,1 The temporal features obtained after inputting into the second BiLSTM network in this branch are: Subsequent use represent

[0089] Multi-scale spatial features extracted by the Inception module Time-series features extracted by the BiLSTM module after flattening by the flattening layer Feature fusion is performed to obtain the spatiotemporal features f of electricity price-related variable data. v Then, the information at different positions in the sequence is associated and interacted with each other through the self-attention mechanism, so as to learn the dependencies in the sequence more comprehensively and pay attention to the important information in the sequence. The calculation process is shown in equations (9) and (10).

[0090] Randomly initialize 4 sets of weight matrices The spatiotemporal characteristics f of electricity price-related variable data v Multiplying by these four sets of weight matrices yields four sets of Q. i K i V i (i∈0,1,2,3); matrix Q i With K i The inner product is normalized using the softmax function to obtain the correlation weight matrix, which is then combined with matrix V. i We perform a weighted summation to obtain the weight coefficients Z; finally, we use Z and the initialization matrix W. 0 Multiplying them together yields the spatiotemporal feature information f that contains the attention mechanism. vSA .

[0091] Finally, the attention feature information obtained from the input branch of the electricity price decomposition matrix and the input branch of the electricity price related variables are fused together, and the predicted electricity price value Y is output after passing through a fully connected layer.

[0092] Step 9: The test set data is input into the trained MBI-IPMDA-PBISA deep learning model to obtain the future electricity price prediction value.

[0093] The embodiment selects variables related to electricity price by using the maximum mutual information coefficient to assist in predicting future electricity price data. Secondly, the original electricity price signal is decomposed using a decomposition algorithm. Then, the electricity price sub-components and electricity price related variables obtained by decomposition are input into the MBI-IPMDA-PBISA deep learning model to predict future electricity prices. In this way, the prediction model can fully learn the complex correlation between electricity price sub-components, electricity price related variables and future electricity prices, and avoid the error accumulation problem caused by separately predicting electricity price sub-components and then superimposing them to obtain future electricity prices. The problem of low prediction accuracy of existing attention mechanisms is solved, and the prediction accuracy is improved.

[0094] Embodiment 2: Taking a DK1 region provided by a certain power exchange as an example:

[0095] Step 1: Collect electricity price data and 27 related variable data such as power generation, wind power generation, electricity consumption, and internal and external electricity trading volume in a certain DK1 region provided by a certain power exchange from December 22, 2021 to December 21, 2022, a total of 365 days.

[0096] Step 2: The missing values in all data are filled using the sliding average window filling method, and the sliding window length is set to 7, so the filled data value is equal to the average of the sum of the first three data points and the last three data points. In order to intuitively show the comparison effect before and after filling the missing values, the DK1 electricity price data on March 27, 2022 is selected and the missing values are filled using the sliding window filling method. The data change curve before and after filling is shown in FIG. 3.

[0097] As shown in FIG. 3, there is a missing value at 03:00 on March 27, 2022, indicating that the DK1 region has missing data in the 02:00-03:00 time period. The sliding window filling method is used to fill the data, and the sliding window is 7, so the filled data value is equal to the average of the sum of the first three data points and the last three data points. The data after sliding window filling is more complete, eliminating the problem of affecting the entire experimental research process due to a missing value.

[0098] Step 3: Normalize all data to unify the dimension size of electricity price and electricity price related variables.

[0099] The Min-Max standardization is used to normalize the electricity price and its related variables after abnormal value processing. Assuming that a vector is X, the Min-Max standardization can be represented as

[0100] wherein: X norm is the value after Min-Max standardization processing, X i is the i-th value of X, X max is the maximum value of X, X min is the maximum value of X. The normalized processing of certain electricity price and its related variable data is shown in FIG. 4. Since there are too many related variables of electricity price, FIG. 4 only shows part of the variables.

[0101] Step 4: Calculate the maximum mutual information coefficient between the electricity price and the electricity price related variables, and the result is shown in FIG. 5.

[0102] According to the calculation result of the maximum mutual information coefficient between the electricity price and the electricity price related variables, finally, 10 variables such as DK2 electricity price, DK1 wind power generation, DK2 wind power generation, DK1 wind power generation prediction value, DK2 wind power generation prediction value, DK power transaction volume, DK and DE power transaction volume, DK and NO power transaction volume, DK1 and DE transaction volume, and DK1 and NL power transaction volume are selected as the electricity price related variables to assist in predicting DK1 electricity price data.

[0103] Step 5: Use the swarm decomposition (SWD) algorithm to decompose the electricity price signal to obtain electricity price sub-components IMF1-IMF10, as shown in FIG. 6.

[0104] Step 6: Divide the electricity price sub-components and the electricity price related variables into model input and output, and use the sliding window mechanism to construct the input and output of the model. The specific construction process is shown in FIG. 7.

[0105] In the figure, n is the length of a certain electricity price data sequence, T ∈ n, and the input data length of the model is set to t. For example, the data in the T+1-T+t time period is input, and the future electricity price data in T+t+1-T+t+24 is predicted, and then the window is slid by 24 steps to obtain the next group of input and output data. In this example, t is set to 48.

[0106] Step 7: Use the last 30 days of data in the 365 days from December 22, 2021 to December 21, 2022 as the test set to test the trained deep learning model, and use the remaining data as the training set to train the deep learning model.

[0107] Step 8: Input the training set data into the MBI-IPMDA-PBISA deep learning model to train the model.

[0108] Step 9: Input the test set data into the trained MBI-IPMDA-PBISA deep learning model to obtain the future electricity price prediction value.

[0109] In order to verify the superiority of the PMDA proposed in the application in the MBI-IPMDA-PBISA prediction model, under the same data preprocessing method, the SE-Net, CBAM and CA are respectively used to replace the PMDA to obtain three prediction models with different attention mechanisms, i.e., MBI-ISE-PBISA, MBI-ICBAM-PBISA and MBI-ICA-PBISA, for day-ahead electricity price prediction, and the experimental error results are compared.

[0110] In order to uniformly evaluate the experimental results and analyze the prediction performance of different models, the mean absolute error (e MAE ), the mean absolute percentage error (e MAPE ), the symmetric mean absolute percentage error (e sMAPE ), the root mean square error (e RMSE) and the determination coefficient (R 2 ) are selected as the performance evaluation functions, and the calculation formulas are as follows:

[0111] In the formulas, n represents the length of the electricity price data sequence to be predicted. P i represents the real value of the electricity price; represents the predicted value of the electricity price; represents the average value of the real value of the predicted day electricity price.

[0112] The prediction results are shown in Table 1 and Fig. 8.

[0113] Table 1: Day-ahead electricity price prediction error of different attention mechanism models in DK1 region

[0114] As shown in Table 1, the MBI-IPMDA-PBISA model with the parallel multi-dimensional attention mechanism proposed in the application has the best prediction effect in the day-ahead electricity price prediction experiment, and the e MAE , e MAPE , e sMAPE , e RMSE and R 2 are 11.6172 (EUR / MWh), 0.0438, 0.0435, 16.1382 (EUR / MWh) and 0.9818, respectively. Compared with the e MAE of the MBI-ISE-PBISA, MBI-ICBAM-PBISA and MBI-ICA-PBISA prediction models in the day-ahead electricity price prediction experiment, the e MAPE are reduced by 21.34%, 12.30% and 5.93%, respectively.decreased by 18.44%, 6.81% and 4.37% respectively, e sMAPE decreased by 22.04%, 10.12% and 2.25% respectively, e RMSE decreased by 18.99%, 11.44% and 4.19% respectively, R 2 increased by 0.0100, 0.0054 and 0.0020 respectively.

[0115] In order to more intuitively show the day-ahead electricity price prediction error of the prediction model with each attention mechanism, each error result is placed in a radar chart for display, and the specific effect is shown in FIG. 9.

[0116] The brown line in FIG. 9 represents the prediction error of the MBI-ISE-PBISA model, the blue line represents the prediction error of the MBI-ICBAM-PBISA model, the purple line represents the prediction error of the MBI-ICA-PBISA model, and the red line represents the prediction error of the MBI-IPMDA-PBISA model proposed in the application. Wherein e MAE ranges from 0 to 15 (EUR / MWh), e MAPE ranges from 0 to 0.1, e sMAPE ranges from 0 to 0.1, e RMSE ranges from 0 to 20 (EUR / MWh), R 2 ranges from 0.9 to 1. It can be seen from FIG. 9 that the prediction error of the MBI-ISE-PBISA prediction model is the largest, and the determination coefficient (R 2 ) is the smallest, indicating that the model with channel attention mechanism has the worst prediction performance, the prediction error of the MBI-ICBAM-PBISA and MBI-ICA-PBISA prediction models is slightly smaller, indicating that the prediction performance of the model with CBAM and CA is slightly better than that of the MBI-ISE-PBISA model, and the prediction error of the MBI-IPMDA-PBISA model with PMDA proposed in the application is the smallest, and R 2 is the highest, indicating that it has the best fitting effect and the best model prediction performance.

[0117] The above specific embodiments are the preferred embodiments of the application, and do not limit the specific implementation range of the application. The scope of the application includes but is not limited to the above specific embodiments. Any equivalent changes made according to the shape, structure and method of the application are within the protection scope of the application.

Claims

1. A day-ahead electricity price forecasting method based on parallel multi-dimensional attention mechanism, characterized in that, The method comprises the following steps: S1, collecting electricity price data and electricity price related variables, pre-processing the collected data, and constructing an electricity price prediction data set; S2, based on the electricity price prediction data set, calculating the maximum mutual information coefficient between the electricity price data and the electricity price related variables, and screening out the electricity price related variables having certain correlation with the electricity price through the maximum mutual information coefficient; S3, decomposing the electricity price data by using a decomposition algorithm to obtain electricity price sub-components; S4, constructing an MBI-IPMDA-PBISA deep learning model; S5, dividing the electricity price sub-components and the electricity price related variables screened out in S2 into input and output to obtain a data set sample, and merging the electricity price sub-components and the electricity price related variables in each sample to obtain an electricity price sub-component input matrix and an electricity price related variable input matrix; S6, dividing the data set sample into a training set and a test set, inputting the training set into the MBI-IPMDA-PBISA deep learning model for training, and then inputting the test set into the trained MBI-IPMDA-PBISA deep learning model for testing to predict future electricity prices.

2. The day-ahead electricity price forecasting method based on parallel multi-dimensional attention mechanism according to claim 1, characterized in that, The data preprocessing comprises processing abnormal values in the electricity price prediction data set and standardizing the electricity price prediction data set.

3. The day-ahead electricity price forecasting method based on parallel multi-dimensional attention mechanism according to claim 2, characterized in that, The processing of the abnormal values comprises: for data missing values, using a sliding window filling method for supplement; for the same time node and repeated data values, directly using one of the data, and for the same time node and different data values, taking the average of the two data for subsequent calculation; for data abnormal values, using 3δ criterion or Grubbs test method to test the electricity price data and the electricity price related variable data, and if it is determined that a point is an abnormal value data, deleting it and then treating it as a data missing value. 4.The parallel multi-dimensional attention mechanism based day-ahead electricity price forecasting method according to claim 1, wherein, The construction of the MBI-IPMDA-PBISA deep learning model comprises: Sa, for the electricity price sub-component input branch, using 2 layers of Inception modules in series to extract multi-scale spatial features of the electricity price sub-components, obtaining a plurality of multi-scale spatial feature matrices, and stacking the plurality of multi-scale spatial feature matrices in the third dimension to obtain multi-scale spatial features of the electricity price sub-components; Sb, adjusting the attention weight on each dimension in the multi-scale spatial features of the electricity price sub-components by using a series of parallel multi-dimensional attention mechanisms; Sc, for the electricity price related variable input branch, using an Inception module and a BiLSTM module in parallel to simultaneously extract spatial features and time sequence features of the electricity price related variables, and fusing the spatial features and the time sequence features to obtain space-time features of the electricity price related variables; Sd, adjusting the internal attention weight in the space-time features of the electricity price related variables by using a series of self-attention mechanisms; Se, fusing the multi-scale spatial features of the electricity price sub-components with the adjusted attention weight on each dimension and the space-time features of the electricity price related variables with the adjusted internal attention weight, and further integrating the fused features through a full connection layer and a regression layer to obtain the output of the MBI-IPMDA-PBISA deep learning model.

5. The day-ahead electricity price forecasting method based on parallel multi-dimensional attention mechanism according to claim 4, characterized in that, The Sa includes: inputting the electricity price sub-component into the first Inception module of the branch to calculate, and the second Inception module under the branch and the Inception module in the electricity price related variable input branch are calculated in the same way, and the multi-scale space feature matrix extracted by the second Inception module of the electricity price sub-component input branch is input into the series-parallel multi-dimensional attention mechanism; wherein the calculation process of the first Inception module of the electricity price sub-component input branch is as follows: In the formula: F p is the Danish price decomposition matrix, a weight matrix for an i-th convolution kernel of the 1x1 convolution branch; weight matrix of the first layer 3x3, 5x5, 7x7 convolution i-th convolution kernel, respectively; weight matrix of the i-th convolution kernel of the second layer 3x3, 5x5, 7x7 convolution respectively; bias for the i-th convolution kernel of the 1 x 1 convolution; bias of the i-th convolution kernel of the first layer 3x3, 5x5, 7x7 convolution respectively; bias of the i-th convolution kernel of the second layer 3x3, 5x5, 7x7 convolution respectively; output the feature for the 1x1 convolution branch; output features of the first layer 3x3, 5x5, 7x7 convolution, respectively; respectively, the output features of the second layer 3x3, 5x5, 7x7 convolution; δ r is a Relu activation function; the feature matrix extracted by the four convolution branches is connected by Concat to obtain the output multi-scale spatial feature matrix Similarly, the multi-scale spatial feature matrix output by the second Inception module in the electricity price sub-component input branch and the Inception module in the electricity price related variable input branch is and 6. The day-ahead electricity price forecasting method based on parallel multi-dimensional attention mechanism according to claim 4, characterized in that, The Sb comprises: Sb1, global max pooling and global average pooling on multi-scale spatial feature matrix The spatial information of the feature map is compressed and aggregated to obtain a feature map with an output size of Cx1x1. Two layers of 1x1 convolution are set to scale learning and obtain the attention weight in the first dimension after being activated by the Sigmoid function. The calculation process is as follows: In the formula: GMP, GAP are global maximum pooling and global average pooling operations; for vector dot product; and weights of the first layer and the second layer in the C dimension for the i-th 1x1 convolution kernel; and bias for the i-th 1x1 convolution kernel of the first layer and the second layer; For global max pooling and global average pooling dot product processed feature output; and f CA is the output after feature learning by the 1x1 convolution of the first layer and the second layer, and σ is a sigmoid activation function. Sb2, the vertical direction spatial information and the horizontal direction spatial information of the multi-scale spatial feature matrix are respectively compressed and aggregated using two one-dimensional maximum pooling and one-dimensional average pooling, to obtain feature maps with output of CxHx1 and Cx1xW, and the calculation process is as shown in the following formula: In the formulae: and For output of the jth channel in the H dimension after average pooling and max pooling; and For The output of the jth channel in the W dimension after average pooling and maximum pooling; τ represents the maximum pooling and average pooling operation with different spatial ranges for each channel; Two layers of 1x1 convolution are set to scale learning and get the attention weights in vertical and horizontal directions after Sigmoid function activation, and the calculation process is shown in the following formula: In the formulae: and weights of the first layer and the second layer in the i-th 1x1 convolution kernel in the H dimension; and a bias of the i-th 1x1 convolution kernel in the H dimension for the first layer and the second layer; and weights of the i-th 1x1 convolution kernel in the W dimension for the first layer and the second layer; and bias of the i-th 1x1 convolution kernel in the first layer and the second layer in the W dimension; and f HA is an output of learning the H-dimension feature through the first layer and the second layer 1x1 convolution; and f WA is the output after learning the W-dimension feature through the 1x1 convolution of the first layer and the second layer; Sb3, multiply the attention weight obtained in different dimensions with the original multi-scale spatial feature matrix to obtain a multi-dimensional attention feature matrix f DA As shown in the following formula:

7. The day-ahead electricity price forecasting method based on parallel multi-dimensional attention mechanism according to claim 4, characterized in that, The Sc includes: For the electricity price related variable input branch, an Inception and BiLSTM parallel structure is adopted to simultaneously extract the spatial features and the time sequence features of the electricity price related variables; the electricity price related variable input matrix is transformed into a 1-dimensional long vector by a flattening layer and is input into the BiLSTM module, the connection between the historical data and the future data is strengthened through the forward and backward propagation rules, and the time sequence features of the electricity price related variables are extracted, and the calculation process is as shown in the following formula; In the formulae: and respectively the output of the forward hidden layer and the backward hidden layer at time t; and input layer to the forward hidden layer and the backward hidden layer, respectively; and are the weight matrices of the output mapping from the previous time output of the forward hidden layer and the backward hidden layer to the current time, respectively; b (1) and b (2) are the biases of the forward hidden layer and the backward hidden layer, respectively; W1and W2are the weight matrices of the output mapping from the output of the forward hidden layer and the backward hidden layer to the output layer, respectively; δ t is the Tanh activation function; for the output of the price-dependent variable matrix after the flattening layer; output of the output layer; similarly, the time sequence feature f v,1 The time sequence feature obtained after inputting the second BiLSTM network of the branch is Subsequently used representative spatial features extracted by the inception module flatly laminated and extracted by BiLSTM module The spatio-temporal features f of the electricity price related variable data are obtained by feature fusion v The calculation process is as shown in the following formula: wherein Indicates concatenation in the first dimension. 8.The parallel multi-dimensional attention mechanism based day-ahead electricity price forecasting method of claim 4, wherein, The Sd includes: By self-attention mechanism, different position information in the space-time characteristics of the electricity price related variable data is associated and interacted with each other, the dependency relationship therein is more comprehensively learned, important information therein is focused, and the calculation process is shown in the following formula: In the formulae: denotes an initialized weight matrix, Q i , K i , V i denote a query vector, a key vector, and a value vector, respectively, d k denotes the dimension of the key vector, Z denotes a weight coefficient, W 0 denotes a weight matrix. 9.The parallel multi-dimensional attention mechanism based day-ahead electricity price forecasting method of claim 1, wherein, The data set sample includes the electricity price sub-component and the electricity price related variable and the future electricity price true value in each sample, and after obtaining the electricity price sub-component input matrix and the electricity price related variable input matrix, the final each sample includes the electricity price sub-component input matrix and the electricity price related variable input matrix and the future electricity price true value.

Images