Residential power load prediction method based on brand new end-to-end

Abstract

The invention discloses a brand new end-to-end-based residential power load prediction method, which comprises the steps of 1) performing feature screening and data preprocessing on a data set by using a Pearson's correlation coefficient method, including feature normalization and one-hot coding; 2) dividing the preprocessed data set into a training set, a test set and a verification set according to 8: 1: 1; 3) extracting features of training set data by adopting a brand new end-to-end network model, learning data distribution and change rules, and constructing a deep learning neural network model; 4) through an Adam algorithm, according to the marble loss function, carrying out back propagation to train model parameters; the training effect of the model parameters is judged according to the loss value of the verification set; and 5) predicting the short-term residential power load by using the optimal deep learning neural network model trained in the step S04, and verifying the prediction precision of the network model.The method has the advantages of high prediction speed, good generalization ability and high prediction result accuracy.

Description

Technical Field

[0001] The present invention relates to the technical field of residential power load forecasting, and in particular to a residential power load forecasting method based on a completely new end-to-end approach. Background Art

[0002] Improving the accuracy and reliability of residential load forecasting will help power plants and power companies obtain the best power generation according to the needs of power users and the power market. In the efficient management and energy-saving operation of smart grids, short-term residential power load forecasting is an indispensable part. At the same time, the rapid growth of population and the development of technology have significantly increased global electricity consumption. Since electricity is consumed at the same time as it is generated by power plants, accurately predicting energy consumption in advance is crucial to ensuring a stable power supply.

[0003] In the past few decades, various short-term residential load forecasting models and methods have been proposed and applied. Traditional time series forecasting methods are mainly divided into two types: statistical methods and traditional machine learning methods; the classic statistical forecasting methods mainly include autoregressive moving average, autoregressive sum moving average model, linear regression method, etc.; for short-term power load forecasting in machine learning, typical methods include support vector regression, fuzzy linear regression, extreme learning machine, etc.; statistical methods can capture the linear relationship of time series, however, existing statistical methods are not suitable for the characteristic dimensions of data sets, data with complex characteristics and rapid dynamic changes, and it is difficult to make real-time predictions of rapidly changing data and achieve reasonable dimensionality reduction compression. Moreover, they only rely on historical load data and cannot make full use of the information provided by additional factors, which hinders the improvement of prediction accuracy. Traditional machine learning methods can effectively fit the implicit nonlinear relationship between historical data and external factors, but there are still some shortcomings that need to be overcome, such as falling into local optimality, relying on manual parameter adjustment, and prone to overfitting. Traditional load forecasting methods such as wavelet analysis and Kalman filtering have short calculation time and fast calculation speed. Classical models are based on statistical calculations and assume linear load sequences. However, with the increasing complexity of power systems, traditional models can no longer meet the requirements. Under conditions involving some nonlinear situations (such as customer behavior), classical methods cannot make accurate predictions due to possible high errors.

[0004] Nowadays, deep learning-based methods are introduced to achieve effective prediction with minimal complexity. Deep learning-based models can effectively learn the complexity, nonlinearity and time dependency of time series load data by stacking the number of hidden layers between the input layer and the output layer. In particular, recurrent neural networks are able to capture complex dependencies in time series data. However, it suffers from the vanishing gradient and exploding gradient problems. In response to the gradient explosion and gradient vanishing problems during the training of recurrent neural models, a gated recurrent neural network represented by LSTM and GRU is proposed to achieve the screening of important historical information through a gating mechanism; although the gated recurrent network is carefully designed to memorize the long-term dependency of the time series, there are still problems of gradient vanishing and gradient explosion when processing too long sequences.

[0005] The quality of short-term residential power load forecasting mainly depends on the model's ability to extract features from historical data. In recent years, deep learning models have been widely used in the field of time series forecasting, including error back propagation neural network (BP), convolutional neural network (CNN), recurrent neural network (RNN), long short-term memory network (LSTM), Transformer, etc. BP is too sensitive to initial weights, which is not conducive to model training. CNN has good local feature extraction capabilities, but ignores the recursiveness of time series. RNN and LSTM have the problem of gradient disappearance or explosion due to chain derivative back propagation. Transformer has excellent performance in computer vision, speech recognition and other fields, but its time and space complexity is high and has a memory bottleneck. Various variants of Transformer have followed, such as Reformer, Informer, Pyraformer, etc., which use sparse attention to reduce complexity, but do not take into account the multi-component nature of time series. Summary of the invention

[0006] The purpose of the present invention is to propose a residential power load forecasting method based on a new end-to-end approach to address the deficiencies of the prior art. On the basis of the encoder in the Transformer, the deep separable convolution and point-by-point convolution structure in the ConvNeXt network are integrated. By introducing the deep separable convolution, the network parameters are reduced, and the time and space information can be better captured, and then all the information is integrated through the point-by-point convolution. By introducing the inverted residual bottleneck structure, more feature information can be effectively retained after the point-by-point convolution performs the operation of "first increasing the dimension and then reducing the dimension", and the prediction model has the advantages of fast prediction speed, good generalization ability and high accuracy of prediction results.

[0007] To achieve the above object, the present invention adopts the following technical solution:

[0008] A new end-to-end residential power load forecasting method includes the following steps:

[0009] Step S01, use the Pearson correlation coefficient method to perform feature screening on the data set and perform data preprocessing, including feature normalization and one-hot encoding;

[0010] Step S02, dividing the preprocessed data set into a training set, a test set and a validation set according to a ratio of 8:1:1;

[0011] Step S03: extract features from the training set data based on a new end-to-end network model, learn its data distribution and change rules, and build a deep learning neural network model;

[0012] Step S04: Use the Adam algorithm to perform back propagation training on the model parameters according to the pinball loss function; and judge the training effect of the model parameters according to the loss value of the validation set;

[0013] Step S05: Use the optimal deep learning neural network model trained in step S04 to predict the short-term residential power load, and compare the predicted results with the actual results through error evaluation indicators to verify the prediction accuracy of the network model.

[0014] Specifically, the data preprocessing in step S01 includes normalizing the average active power load of households and weather, and performing unique hot encoding on seasons and holidays as well as the active power watt-hours of various household electrical equipment; the normalization process records the data of each column of the data set feature in chronological order as X={x1,x2,…,x n}, use Max-Min normalization to map the data to between 0 and 1. The calculation formula is:

[0015]

[0016] In the above formula, x max and x min are the maximum and minimum values ​​of the column data respectively, and the new sequence is recorded as Y={y1,y2,…,y n};

[0017] The one-hot encoding process is as follows: if the data feature is spring, its feature is encoded as [1,0,0,0]; if the data feature is on holidays, its feature is encoded as [1,0]; if the data feature is not on holidays, it is encoded as [0,1]; if the household electrical equipment is in use, its feature is encoded as [1,0]; if it is not in use, it is encoded as [0,1].

[0018] Specifically, in step S03, features are extracted from the training set data based on a new end-to-end network model, wherein the main structure of the new end-to-end network model includes a Transformer sequence-to-sequence structure, an encoder, a TCN network integrated into the encoder, and a ConvNeXt network. The networks are connected by residuals to add the model input data and output data, and the result is used as the final output data of the encoder.

[0019] The standard convolutional layer of the ConvNeXt network adopts a size of D K ×D K ×M×N convolution kernel K is parameterized, where D K is the spatial dimension of the kernel assumed to be square, M is the number of input channels, and N is the number of output channels; the computational cost of standard convolution is: D K ·D K ·M·N·D F ·D F , where the computational cost depends multiplicatively on the number of input channels M, the number of output channels N, and the convolution kernel size D K ×D K and feature map D F ×D F Size; for depth-wise separable convolution, the computational cost is: D K ·D K ·M·D F ·D F +D F ·D F ·N.

[0020] In the above technical solution, the new end-to-end network model retains the advantages of the TCN network structure, namely, dilated causal convolution and residual connection. The causal convolution ensures strict time constraints and performs convolution operations in chronological order, that is, the value at time t depends only on the value of the previous layer of the network at time t and before time t; in addition, in order to learn longer temporal dependencies and avoid the loss of historical data information, dilated convolution is used, which can increase the size of the receptive field without pooling, so that each convolution outputs a wider range of information; at the same time, in order to increase the stability of the network and prevent the number of network layers from causing network degradation, residual connections are added to the network, that is, the input data directly skips the intermediate convolution operation, so that the input data and the output data are added, and the result is used as the final output data of the encoder;

[0021] The advantages of the ConvNeXt network structure are that it can adapt to multi-dimensional input data, better process spatial information, reduce model parameters, and also has good generalization ability.

[0022] Specifically, the deep learning neural network model in step S03 has an encoder-decoder structure, and the encoder includes three modules: an autocorrelation calculation module, a causal convolution module, and a deep separable convolution module; the decoder has only an LSTM module; the autocorrelation calculation module and the causal convolution module are used to determine the period-based correlation by calculating the continuity of sequence autocorrelation and time delay aggregation similarity; the deep separable convolution module is used to extract the spatial correlation between multiple load variables and to better learn the correlation between channels and between spaces; the time series data processed by the three modules of seasonality, trend, and spatial correlation in the encoder will obtain corresponding more complete feature representations, which will be passed to the LSTM module in the decoder for prediction.

[0023] Specifically, the core structure of the LSTM module includes a forget gate, an input gate, and an output gate. The forget gate can decide to discard part of the historical information from the unit state. When the value of the forget gate is 1, the information will be completely retained. Conversely, when it is 0, all historical information will be discarded, that is, the forget gate can selectively forget the historical information; when the input gate and the output gate update the unit state, they can control the information input into the current unit and generate new information to be output; the output data from the encoder is used as the input of the LSTM module in the decoder, and the predicted value is finally output. The LSTM module can effectively capture long-term dependencies in time series prediction, use historical information, and has good adaptability. Therefore, it is widely used in time series prediction and can improve the accuracy of predicted values ​​in short-term load prediction.

[0024] Furthermore, the deep learning neural network model also includes an activation function and a normalization function. The activation function selects a Gaussian error linear unit GELU, and the output of the power model is converted into nonlinearity through the GELU function without changing the shape of the data. The calculation formula of the GELU function is:

[0025] GELU(x)=x*φ(x) (2)

[0026] In the above formula, φ(x) represents the cumulative probability distribution of Gaussian distribution, that is, the definite integral of Gaussian distribution in the interval (-∞,x]);

[0027] The normalization function selects a layer normalization function, which is used to comprehensively consider the input of all dimensions of a layer, calculate the average input value and input variance of the layer, and then use the same normalization operation to transform the input of each dimension. Layer normalization is performed on a single training sample and does not depend on other data. Therefore, it can avoid the problem of batch normalization being affected by the distribution of small batch data. It can be used in small batch scenarios, dynamic network scenarios and RNN, especially in the field of natural language processing; in addition, layer normalization does not need to save the mean and variance of small batch data, saving additional storage space.

[0028] Furthermore, the layer normalization function and activation function are only used once in each module of the deep learning neural network model, which can better retain feature information.

[0029] Compared with the prior art, the present invention has the following beneficial effects:

[0030] In order to improve the reliability of short-term residential power load forecasting, the method of the present invention integrates the deep separable convolution and point-by-point convolution structure in the ConvNeXt network on the basis of the encoder in the Transformer. By introducing the deep separable convolution, the network parameters are reduced, and the time and space information can be better captured, and then all the information is integrated through point-by-point convolution. By introducing the inverted residual bottleneck structure, after the point-by-point convolution performs the operation of "first increasing the dimension and then reducing the dimension", more feature information can be effectively retained, which can make the final prediction result more accurate. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 It is an overall flow chart of a new end-to-end residential power load forecasting method according to the present invention;

[0032] Figure 2 It is a structural diagram of the present invention based on a new end-to-end residential power load prediction model;

[0033] Figure 3 This is the TCN network module structure diagram;

[0034] Figure 4 This is the ConvNeXt network module structure diagram;

[0035] Figure 5 This is the LSTM module structure diagram. DETAILED DESCRIPTION

[0036] In order to facilitate the understanding of the present invention, the present invention will be described more fully below with reference to the relevant drawings. Several embodiments of the present invention are given in the drawings. However, the present invention can be implemented in many different forms and is not limited to the embodiments described herein. On the contrary, the purpose of providing these embodiments is to make the disclosure of the present invention more thorough and comprehensive.

[0037] Example

[0038] like Figure 1 As shown, the present invention provides a residential power load forecasting method based on a new end-to-end system, comprising the following steps:

[0039] Step S01: Use the Pearson correlation coefficient method to perform feature screening on the data set and perform data preprocessing, including feature normalization and one-hot encoding;

[0040] Step S02, divide the preprocessed data set into a training set, a test set, and a validation set according to a ratio of 8:1:1; this step requires setting the batch size to convert the data into small batches to increase the training speed. In this embodiment, the batch size is set to 512;

[0041] Step S03: extract features from the training set data based on a new end-to-end network model, learn its data distribution and change rules, and build a deep learning neural network model;

[0042] Step S04: Use the Adam algorithm to perform back propagation training on the model parameters according to the pinball loss function; and judge the training effect of the model parameters according to the loss value of the validation set;

[0043] Step S05: Use the optimal deep learning neural network model trained in step S04 to predict the short-term residential power load, and compare the predicted results with the actual results through error evaluation indicators to verify the prediction accuracy of the network model.

[0044] Specifically, the data preprocessing in step S01 includes normalizing the average active power load of households and weather, and performing unique hot encoding on seasons and holidays as well as the active power watt-hours of various household electrical equipment; the normalization process records the data of each column of the data set feature in chronological order as X={x1,x2,…,x n}, use Max-Min normalization to map the data to between 0 and 1. The calculation formula is:

[0045]

[0046] In the above formula, x max and x minare the maximum and minimum values ​​of the column data respectively, and the new sequence is recorded as Y={y1,y2,…,y n};

[0047] The one-hot encoding process is as follows: if the data feature is spring, its feature is encoded as [1,0,0,0]; if the data feature is on holidays, its feature is encoded as [1,0]; if the data feature is not on holidays, it is encoded as [0,1]; if the household electrical equipment is in use, its feature is encoded as [1,0]; if it is not in use, it is encoded as [0,1].

[0048] Specifically, in step S03, features are extracted from the training set data based on a new end-to-end network model, such as Figure 2 As shown in FIG. 1 , the main structure of the new end-to-end network model includes a Transformer sequence-to-sequence structure, an encoder, a TCN network integrated into the encoder, and a ConvNeXt network. The networks are connected through residual connections to add the model input data and output data, and the result is used as the final output data of the encoder.

[0049] Furthermore, if Figure 4 As shown, another advantage of the ConvNeXt network over standard convolution is that the computational cost of depthwise separable convolution is lower. Its standard convolution layer uses a size of D K ×D K ×M×N convolution kernel K is parameterized, where D K is the spatial dimension of the kernel assumed to be square, M is the number of input channels, and N is the number of output channels; the computational cost of standard convolution is: D K ·D K ·M·N·D F ·D F , where the computational cost depends multiplicatively on the number of input channels M, the number of output channels N, and the convolution kernel size D K ×D K And feature map D F ×D F Size; for depth-wise separable convolution, the computational cost is: D K ·D K ·M·D F ·D F +D F ·D F N. Therefore, the depth-wise separable convolution can do the same work as the standard convolution with reduced computational cost; the advantage of the ConvNeXt network structure is that it can adapt to multi-dimensional input data, better process spatial information, reduce model parameters, and has good generalization ability. These advantages make it perform well in residential power load forecasting.

[0050] In this embodiment, the new end-to-end network model is improved based on the Transformer architecture. Figure 3 The figure shows the TCN network structure diagram. The model retains the advantages of the TCN network structure, namely, the dilated causal convolution and residual connection. The causal convolution ensures strict time constraints and performs convolution operations in chronological order, that is, the value at time t only depends on the value of the previous layer of the network at time t and before time t; in addition, in order to learn longer temporal dependencies and avoid the loss of historical data information, dilated convolution is used, which can increase the size of the receptive field without pooling, so that each convolution outputs a wider range of information; at the same time, in order to increase the stability of the network and prevent the number of network layers from increasing and causing network degradation, residual connections are added to the network, that is, the input data directly skips the intermediate convolution operation, so that the input data and output data are added, and the result is used as the final output data of the encoder, and the output data is used as the input of the LSTM in the decoder to finally obtain the predicted value.

[0051] Specifically, the deep learning neural network model in step S03 has an encoder-decoder structure, and the encoder includes three modules: an autocorrelation calculation module, a causal convolution module, and a deep separable convolution module; the decoder has only an LSTM module; the autocorrelation calculation module and the causal convolution module are used to determine the period-based correlation by calculating the continuity of sequence autocorrelation and time delay aggregation similarity; the deep separable convolution module is used to extract the spatial correlation between multiple load variables and to better learn the correlation between channels and between spaces; the time series data processed by the three modules of seasonality, trend, and spatial correlation in the encoder will obtain corresponding more complete feature representations, which will be passed to the LSTM module in the decoder for prediction.

[0052] Specifically, Figure 5 As shown in the figure, the core structure of the LSTM module includes a forget gate, an input gate and an output gate. The forget gate can decide to discard part of the historical information from the unit state. When the value of the forget gate is 1, the information will be completely retained. Conversely, when it is 0, all historical information will be discarded, that is, the forget gate can selectively forget the historical information; when the input gate and the output gate update the unit state, they can control the information input into the current unit and generate new information to be output; the output data from the encoder is used as the input of the LSTM module in the decoder, and the predicted value is finally output. The LSTM module can effectively capture long-term dependencies in time series prediction, use historical information, and has good adaptability. Therefore, it is widely used in time series prediction and can improve the accuracy of predicted values ​​in short-term load prediction.

[0053] Furthermore, the deep learning neural network model also includes an activation function and a normalization function. The activation function selects a Gaussian error linear unit GELU, and the output of the power model is converted into nonlinearity through the GELU function without changing the shape of the data. The calculation formula of the GELU function is:

[0054] GELU(x)=x*φ(x) (2)

[0055] In the above formula, φ(x) represents the cumulative probability distribution of Gaussian distribution, that is, the definite integral of Gaussian distribution in the interval (-∞,x]);

[0056] The normalization function selects a layer normalization function, which is used to comprehensively consider the input of all dimensions of a layer, calculate the average input value and input variance of the layer, and then use the same normalization operation to transform the input of each dimension. Layer normalization is performed on a single training sample and does not depend on other data. Therefore, it can avoid the problem of batch normalization being affected by the distribution of small batch data. It can be used in small batch scenarios, dynamic network scenarios and RNN, especially in the field of natural language processing; in addition, layer normalization does not need to save the mean and variance of small batch data, saving additional storage space.

[0057] Furthermore, the layer normalization function and activation function are only used once in each module of the deep learning neural network model, which can better retain feature information.

[0058] The performance of the prediction method of the present invention is further evaluated by an example below.

[0059] The data set used for the performance evaluation of the present invention is from the electricity consumption data of a French family for four years. The power load data is selected from 68,245 sets of data from December 16, 2006 to May 12, 2010, with a resolution of 15 minutes. The data set is split into a training set, a validation set, and a test set.

[0060] The error evaluation index evaluation model used in this example is the mean absolute percentage error MAPE, the mean absolute error MAE and the root mean square error RMSE. The specific calculation formulas are as follows:

[0061]

[0062]

[0063] In the above formula, Indicates the predicted value corresponding to the training result of the i-th sample of the prediction model, y i It represents the true value corresponding to the training result of the i-th sample of the prediction model, and n is the total number of samples in the prediction model.

[0064] The above three error evaluation models are used to evaluate the prediction performance of the method of the present invention.

[0065] The comparison of the prediction effect of the prediction model of the present invention and other models on the above data set is shown in Table 1 below.

[0066] Table 1. Comparison of prediction model results

[0067]

[0068] By comparing with other prediction models, it can be seen that the prediction method proposed by the present invention has higher accuracy than other models, and its efficiency is also improved compared with the traditional Transformer model, and can obtain accurate results in residential load prediction.

[0069] It should be understood that the above description of the preferred embodiment is relatively detailed and cannot be regarded as limiting the scope of patent protection of the present invention. Under the enlightenment of the present invention, ordinary technicians in this field can also make substitutions or modifications without departing from the scope of protection of the claims of the present invention, which all fall within the scope of protection of the present invention. The scope of protection requested for the present invention shall be based on the attached claims.

Claims

1. A new end-to-end residential power load forecasting method, characterized in that: The following steps are involved: Step S01, use the Pearson correlation coefficient method to perform feature screening on the data set and perform data preprocessing, including feature normalization and one-hot encoding; Step S02, dividing the preprocessed data set into a training set, a test set and a validation set according to a ratio of 8:1:1; Step S03: extract features from the training set data based on a new end-to-end network model, learn its data distribution and change rules, and build a deep learning neural network model; Step S04: Use the Adam algorithm to perform back propagation training on the model parameters according to the pinball loss function; and judge the training effect of the model parameters according to the loss value of the validation set; Step S05: Use the optimal deep learning neural network model trained in step S04 to predict the short-term residential power load, and compare the predicted results with the actual results through error evaluation indicators to verify the prediction accuracy of the network model.

2. A residential power load forecasting method based on a new end-to-end approach according to claim 1, characterized in that: The data preprocessing in step S01 includes normalizing the average active power load of households and weather, and performing unique hot encoding on seasons and holidays as well as the active power watt-hours of various household electrical equipment; the normalization process records the data of each column of the data set feature in chronological order as X={x1,x2,…,x n }, use Max-Min normalization to map the data to between 0 and 1. The calculation formula is: In the above formula, x max and x min are the maximum and minimum values ​​of the column data respectively, and the new sequence is recorded as Y={y1,y2,…,y n }; The one-hot encoding process is as follows: if the data feature is spring, its feature is encoded as [1,0,0,0]; if the data feature is on holidays, its feature is encoded as [1,0]; if the data feature is not on holidays, it is encoded as [0,1]; if the household electrical equipment is in use, its feature is encoded as [1,0]; if it is not in use, it is encoded as [0,1].

3. A residential power load forecasting method based on a new end-to-end approach according to claim 1, characterized in that: In step S03, features are extracted from the training set data based on a new end-to-end network model. The main structure of the new end-to-end network model includes a Transformer sequence-to-sequence structure, an encoder, a TCN network integrated into the encoder, and a ConvNeXt network. The networks are connected by residuals to add the model input data and output data, and the result is used as the final output data of the encoder. The standard convolutional layer of the ConvNeXt network adopts a size of D K ×D K ×M×N convolution kernel K is parameterized, where D K is the spatial dimension of the kernel assumed to be square, M is the number of input channels, and N is the number of output channels; the computational cost of standard convolution is: D K ·D K ·M·N·D F ·D F , where the computational cost depends multiplicatively on the number of input channels M, the number of output channels N, and the convolution kernel size D K ×D K and feature map D F ×D F Size; for depth-wise separable convolution, the computational cost is: D K ·D K ·M·D F ·D F +D F ·D F ·N.

4. A residential power load forecasting method based on a new end-to-end approach according to claim 1, characterized in that: The deep learning neural network model in step S03 has an encoder-decoder structure, wherein the encoder includes three modules: an autocorrelation calculation module, a causal convolution module, and a depth-separable convolution module; the decoder includes only an LSTM module; the autocorrelation calculation module and the causal convolution module are used to determine the period-based correlation by calculating the continuity of sequence autocorrelation and time-delayed aggregation similarity; The depth-separable convolution module is used to extract the spatial correlation between multiple load variables and to better learn the correlation between channels and spaces; The time series data processed by the three modules of seasonality, trend and spatial correlation in the encoder will obtain corresponding more complete feature representations, which will be passed to the LSTM module in the decoder for prediction.

5. A residential power load forecasting method based on a new end-to-end approach according to claim 4, characterized in that: The core structure of the LSTM module includes a forget gate, an input gate and an output gate. The forget gate can decide to discard part of the historical information from the unit state. When the value of the forget gate is 1, the information will be completely retained. Conversely, when it is 0, all historical information will be discarded, that is, the forget gate can selectively forget the historical information. When updating the unit state, the input gate and the output gate can control the information input into the current unit and generate new information to be output; The output data from the encoder is used as the input of the LSTM module in the decoder, and finally the predicted value is output.

6. A residential power load forecasting method based on a new end-to-end approach according to claim 1, characterized in that: The deep learning neural network model also includes an activation function and a normalization function. The activation function selects a Gaussian error linear unit GELU, and the output of the power model is converted into nonlinearity through the GELU function without changing the shape of the data. The calculation formula of the GELU function is: GELU(x)=x*φ(x) (2) In the above formula, φ(x) represents the cumulative probability distribution of Gaussian distribution, that is, the definite integral of Gaussian distribution in the interval (-∞,x]); The normalization function selects a layer normalization function, which is used to comprehensively consider the inputs of all dimensions of a layer, calculate the average input value and input variance of the layer, and then use the same normalization operation to transform the inputs of each dimension.

7. A new end-to-end residential power load forecasting method according to claim 6, characterized in that: The layer normalization function and activation function are used only once in each module of the deep learning neural network model.

Images