Short-term power load prediction method combining improved empirical mode decomposition and bidirectional long short-term memory network

Abstract

The invention relates to the technical field of short-term power load prediction, in particular to a short-term power load prediction method combining improved empirical mode decomposition and a bidirectional long short-term memory network, which comprises the following steps: preprocessing historical load data; decomposing the load data into a plurality of intrinsic mode functions by adopting a complete adaptive noise integrated empirical mode decomposition algorithm, and adaptively adding Gaussian white noise to suppress a mode aliasing problem; constructing a prediction model based on a bidirectional long-short-term memory network and a self-attention mechanism for each component; extracting features by an input layer through a sliding window; the training layer adopts a bidirectional long short-term memory network to extract time sequence hidden features, and a key time point is focused through weighting of a self-attention mechanism; the output layer generates a component prediction value; and superposing all component prediction results, and reconstructing to obtain an accurate load prediction value of the to-be-predicted day. According to the method, the prediction precision and stability are effectively improved by improving signal decomposition and deep learning model fusion.

Description

Technical Field

[0001] The present invention relates to the technical field of short-term power load forecasting, and in particular to a short-term power load forecasting method combining improved empirical mode decomposition and a bidirectional long short-term memory network. Background Art

[0002] With the development of intelligent power systems, the importance of short-term load forecasting for safe grid dispatch and energy optimization management has become increasingly prominent. Traditional forecasting methods are unable to effectively cope with challenges such as strong non-stationarity, noise interference and complex time series dependencies of load data. Empirical mode decomposition is prone to modal aliasing when processing actual load signals, resulting in the intrinsic mode function containing mixed scale characteristics. A single forecasting model (such as the standard LSTM) is unable to fully exploit the long-term and short-term dependencies and key time point characteristics of load data.

[0003] The Chinese invention patent application with publication number CN118052330A discloses a short-term power load forecasting method and system based on ICEEMDAN-LSTM, including obtaining historical power load data, constructing a historical data set, and preprocessing the historical data set; decomposing the preprocessed historical data set through ICEEMDAN to obtain modal components; inputting the modal components into a pre-established LSTM model for training, and performing power load forecasting to obtain prediction results; the invention application predicts each IMFs component through LSTM, and accumulates its prediction results to obtain the final prediction result, which can greatly improve the accuracy of the power load forecast result, greatly improves the accuracy of load forecasting, and compared with direct prediction, can identify the characteristics of each frequency load component, and is suitable for the situation where the load fluctuation is large under the current distributed energy generation background.

[0004] However, power load data has inherent characteristics such as non-stationarity, noise interference and complex time series dependence. The empirical mode decomposition in existing technologies may produce modal aliasing when processing multi-scale signals, affecting the quality of the intrinsic mode function; at the same time, there is room for improvement in the ability of a single prediction model to capture long-term time series characteristics and key time point information; the coordinated optimization of signal decomposition quality and model feature extraction capabilities remains the key direction of technological development in this field. Summary of the Invention

[0005] The purpose of the present invention is to address the problems existing in the background technology and propose a short-term power load forecasting method combining improved empirical mode decomposition and bidirectional long short-term memory network.

[0006] The technical solution of the present invention is a short-term power load forecasting method combining improved empirical mode decomposition and bidirectional long short-term memory network, which includes the following specific implementation steps:

[0007] S1, preprocessing historical load data, including missing value filling and outlier correction;

[0008] S2, decomposing the preprocessed historical load data into a plurality of intrinsic mode function components by using a complete adaptive noise integrated empirical mode decomposition algorithm;

[0009] S3, for each intrinsic mode function component, constructing a prediction model based on a self-attention mechanism and a bidirectional long short-term memory network to generate a component prediction value;

[0010] S4, superimposing all component prediction values to obtain a short-term power load prediction result of the day to be predicted.

[0011] Preferably, the missing value filling adopts a mean filling method, and the specific formula is:

[0012] wherein, represents the load mean value of the nth day; x n,i represents the load value of the ith moment of the nth day; I is 48, that is, 48 data collected in 24 hours; x defect represents the filled missing value.

[0013] Preferably, the outlier correction includes:

[0014]

[0015] wherein, x i represents the mean value of the ith moment in a day; σ i represents the standard deviation of the ith moment, and ε represents a set abnormal threshold value;

[0016] If the absolute value of the difference between the load value x n,i and is greater than 3σ i ε, it is judged as an outlier, and the outlier is corrected, and the correction formula is:

[0017]

[0018] wherein, x n,i ' represents the corrected value of the ith moment of the nth day; x m,i represents the data value of the same kind of day at the ith moment; and α and β are weight coefficients set.

[0019] Preferably, the complete adaptive noise integrated empirical mode decomposition algorithm includes:

[0020] S21, adding I times Gaussian white noise to the historical load data y(t): y i (t)=y(t)+ε'N i (t);

[0021] Among them, N i (t) represents a Gaussian white noise signal that satisfies the standard normal distribution, i = 1, 2, ..., I; ε' represents an adaptive function; y i (t) represents the new data obtained after adding white noise;

[0022] S22, for the above y i (t) Perform empirical mode decomposition to obtain I first-order modal components, and average them to obtain the first eigenmode function C1'(t):

[0023]

[0024] Among them, EMD(y i (t)) represents the new data y obtained after adding white noise i (t) performing an empirical mode decomposition operation, i.e., EMD decomposition; Represents the new data y obtained after adding white noise i (t) The first-order modal component obtained after empirical mode decomposition; r i (t) represents the new data y obtained after adding white noise i (t) is the single decomposition residual obtained after empirical mode decomposition; C1'(t) represents the first intrinsic mode function;

[0025] S23. For the kth decomposition, first find the remainder r after k-1 decompositions. k-1 (t), and then find the kth eigenmode function according to the following method:

[0026]

[0027] Among them, r k-1 (t) represents the total residual after the first k-1 adaptive noise integrated EMD decompositions; C k '(t) represents the intrinsic mode function obtained by the kth decomposition; EMD1(·) represents the first-order mode component obtained by the empirical mode decomposition; EMD k-1 (·) represents the k-1th order modal component obtained by empirical mode decomposition;

[0028] S24. When the number of extreme points of the residual is ≤ 2, the decomposition is terminated, and K intrinsic mode functions and residual r after K decompositions are obtained. K (t), and based on this we can get the historical load data y(t):

[0029]

[0030] Preferably, the condition for terminating the decomposition is: if the residue r after the kth decomposition is kIf the number of extreme points of (t) is less than or equal to 2, the decomposition process is stopped.

[0031] Preferably, the prediction model based on the self-attention mechanism and the bidirectional long short-term memory network includes:

[0032] Input layer: The input variable is the load component data mentioned above, the number of components is n, and the load x is extracted using a sliding window to form the input feature vector X:

[0033]

[0034] in, is the feature vector extracted from the jth sliding window of the i-th load component; t is the prediction time; h is the prediction horizon; lag is the number of consecutive lagged loads; m is the width of the sliding window, the window sliding step is 1, and the number of sliding windows T = lag + 1-m;

[0035] Model training layer: It consists of a bidirectional long short-term memory network layer, a self-attention mechanism layer, and a second bidirectional long short-term memory network layer, which is used to extract temporal features and weight key time points;

[0036] Output layer: The output layer consists of a fully connected layer, which is used to obtain the predicted value of the load component and output the predicted vector of the load component

[0037] Among them, Y (i) represents the prediction result vector of the i-th load component; It represents the predicted value of the i-th load component at the start time of the prediction time window, t is the target time point of the prediction, and h is the prediction range.

[0038] Preferably, the model training layer includes:

[0039] Bidirectional long short-term memory network layer, which extracts hidden features from the load component data. The hidden feature vector can be expressed as: h t =[h t1 ,h t2 ,…,h tm ], t∈[1,T];

[0040] Among them, h t is the hidden feature vector; m and T are the dimension of the hidden feature and the number of time steps respectively;

[0041] The attention layer passes the extracted hidden features to the self-attention mechanism layer, which weights the hidden features through the self-attention mechanism. The process of weighting the hidden features includes:

[0042] Calculate the score of hidden features of each dimension at different time steps:

[0043] s j,d =f sco (W j,d [h 1,d ,h 2,d ,…,h T,d ]),d=1,…,m,j=1,…,T;

[0044] Among them, h j,d represents the hidden features of d dimensions at the jth time step; W j,d Represents the weight vector adjusted during training; function f sco It is implemented through a fully connected layer, whose number of neurons is equal to the number of time steps T;

[0045] Use the softmax function to normalize the scores of a specific dimension so that the sum of the scores at different time steps is 1:

[0046]

[0047] Among them, s j,d 'Indicates the score after normalizing the feature dimension score using the softmax function;

[0048] The scores of all dimensions at the same time step are averaged to obtain the attention weight at each time step. The hidden layer features of the corresponding time step are then multiplied by the attention weight, and the resulting new feature matrix is ​​used as the input of the subsequent layer:

[0049]

[0050] Among them, α j represents the self-attention weight score at the jth time step; h j represents the hidden feature vector at the jth time step; h j wt represents the weighted hidden feature vector at the jth time step;

[0051] The second bidirectional long short-term memory network layer is responsible for extracting more abstract features from the weighted hidden features passed above.

[0052] Preferably, the short-term power load forecast result for the forecast day is:

[0053]

[0054] Among them, Y is the short-term power load forecast result of the forecast day; Y (i) is the prediction result vector of the i-th load component.

[0055] Compared with the prior art, the above technical solution of the present invention has the following beneficial technical effects:

[0056] The present invention designs a short-term power load forecasting method that combines improved empirical mode decomposition and bidirectional long short-term memory network. By combining the improved empirical mode decomposition and bidirectional long short-term memory network, the accuracy and stability of short-term power load forecasting are significantly improved; the improved complete adaptive noise integrated empirical mode decomposition algorithm effectively suppresses the modal aliasing problem in traditional empirical mode decomposition, and decomposes the load data into a more stable intrinsic mode function by adaptively adding Gaussian white noise, laying a high-quality data foundation for subsequent predictions; the prediction model constructed for each component after decomposition integrates the bidirectional long short-term memory network and the self-attention mechanism. The former fully explores the front-end dependency relationship of the time series, and the latter adaptively weights the characteristics of key time points, enhancing the ability to capture complex load changes; finally, high-precision load forecasting is achieved through the superposition and reconstruction of component prediction values;

[0057] The present invention has significant advantages in reducing noise interference, overcoming modal aliasing, and improving timing modeling capabilities, and can be widely used in the fields of power system scheduling and energy management. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 A flow chart of a method for short-term power load forecasting proposed by the present invention that combines improved empirical mode decomposition and bidirectional long short-term memory network;

[0059] Figure 2 This is a flow chart of a complete adaptive noise integrated empirical mode decomposition proposed by the present invention;

[0060] Figure 3 This is a structural diagram of a prediction model based on self-attention mechanism and bidirectional long short-term memory network proposed by the present invention;

[0061] Reference numerals: IMF1, IMF2, ..., IMF K are the K intrinsic mode function components obtained by the complete empirical mode decomposition of adaptive noise; r K represents the residual after K-times decomposition; BiLSTM-SAM is a prediction model based on self-attention mechanism and bidirectional long short-term memory network; Y1, Y2, ..., Y K 、Y K+1 is the prediction result vector; EMD is the empirical mode decomposition operation; IMFk1, IMFk2, ..., IMFk I is the intrinsic mode function component obtained by the complete empirical mode decomposition of adaptive noise; BiLSTM is a bidirectional long short-term memory network. DETAILED DESCRIPTION

[0062] In the first embodiment, the present invention proposes a short-term power load forecasting method combining improved empirical mode decomposition and bidirectional long short-term memory network, such as Figure 1 The specific implementation steps include the following:

[0063] S1, collecting power data in real time through intelligent power equipment sensors, and preprocessing the collected data to construct a historical load data set;

[0064] Specifically, the data preprocessing includes:

[0065] S11, missing value filling: due to sensor failure and human error, etc., missing values may occur during the collection process, which are filled by analyzing the data values near the missing values, and the following mean filling method is used to fill them:

[0066] wherein, represents the load mean value of the nth day; x n,i represents the load value of the ith moment of the nth day; I is 48, i.e., 48 data collected within 24 hours; x defect represents the filled missing value;

[0067] S12, abnormal value correction: abnormal values in the data will affect the accuracy of the prediction, and the detection and correction process of the abnormal values is as follows:

[0068] The standard deviation is used to detect abnormal values, and the judgment criteria are as follows:

[0069]

[0070] wherein, x i represents the mean value of the ith moment in a day; σ i represents the standard deviation of the ith moment, and ε represents the set abnormal threshold, which is set to 1.2 in this embodiment;

[0071] If the absolute value of the difference between the load value x n,i and is greater than 3σ i ε, it is judged to be an abnormal value, which is corrected, and the correction formula is:

[0072]

[0073] wherein, x n,i ' represents the corrected value of the ith moment of the nth day; x m,i represents the data value of the ith moment of the same day; and α and β are weight coefficients set.

[0074] S2. Decompose the historical load data set into multiple more stable intrinsic mode functions (IMFs) using a complete adaptive noise integrated empirical mode decomposition algorithm. The complete adaptive noise integrated empirical mode decomposition algorithm process is as follows: Figure 2 As shown, specifically:

[0075] S21. Add I-order Gaussian white noise to the historical load data y(t): y i (t)=y(t)+ε'N i (t);

[0076] Among them, N i (t) represents a Gaussian white noise signal that satisfies the standard normal distribution, i = 1, 2, ..., I; ε' represents an adaptive function; y i (t) represents the new data obtained after adding white noise;

[0077] S22, for the above y i (t) Perform empirical mode decomposition to obtain I first-order modal components, and then average them to obtain the first eigenmode function C1'(t):

[0078]

[0079] Among them, EMD(y i (t)) represents the new data y obtained after adding white noise i (t) performing an empirical mode decomposition (EMD) operation; Represents the new data y obtained after adding white noise i (t) The first-order modal component obtained after empirical mode decomposition; r i (t) represents the new data y obtained after adding white noise i (t) is the single decomposition residual obtained after empirical mode decomposition; C1'(t) represents the first intrinsic mode function;

[0080] S23. For the kth (k>1) decomposition, first find the remainder r after the k-1th decomposition. k-1 (t), and then find the kth eigenmode function according to the following method:

[0081]

[0082] Among them, r k-1 (t) represents the total residual after the first k-1 adaptive noise integrated EMD decompositions; C k '(t) represents the intrinsic mode function (IMF) obtained by the k-th decomposition; EMD1(·) represents the first-order mode component obtained by the empirical mode decomposition; EMDk-1 (·) represents the k-1th order modal component obtained by empirical mode decomposition;

[0083] S24, determine the remainder r after the kth decomposition k Is the number of extreme points of (t) greater than 2?

[0084] If it is greater than 2, continue to step S23;

[0085] If it is less than 2, the remainder cannot be decomposed further, and the decomposition process ends;

[0086] After K decompositions, we get K intrinsic mode functions and residual r K (t), and based on this we can get the historical load data y(t):

[0087]

[0088] S3. For each of the load components, a prediction model based on the self-attention mechanism and the bidirectional long short-term memory network is constructed. The structure of the network model is as follows: Figure 3 As shown in Figure 2, the input of the model is the historical load component sequence, and the output of the model is the predicted value of the load component on the day to be predicted. The short-term load forecasting model based on the self-attention mechanism and bidirectional long short-term memory network includes:

[0089] (1) Input layer: The input variable is the load component data mentioned above, the number of components is n, and the load x is extracted using a sliding window to form the input feature vector X:

[0090]

[0091] in, is the feature vector extracted from the jth sliding window of the i-th load component; t is the prediction time; h is the prediction horizon; lag is the number of consecutive lagged loads; m is the width of the sliding window, the window sliding step is 1, and the number of sliding windows T = lag + 1-m;

[0092] (2) Model training layer: The training layer mainly consists of two bidirectional long short-term memory network layers and a self-attention mechanism layer, specifically:

[0093] The first layer is a bidirectional long short-term memory network layer, which extracts hidden features from the load component data. The hidden feature vector can be expressed as: h t =[h t1 ,h t2 ,…,h tm ], t∈[1,T];

[0094] Among them, h t is the hidden feature vector; m and T are the dimension of the hidden feature and the number of time steps respectively;

[0095] The second layer is the attention layer, which passes the extracted hidden features to the self-attention mechanism layer. This layer weights the hidden features through the self-attention mechanism. The process of weighting the hidden features includes:

[0096] Calculate the score of hidden features of each dimension at different time steps:

[0097] s j,d =f sco (W j,d [h 1,d ,h 2,d ,…,h T,d ]),d=1,…,m,j=1,…,T;

[0098] Among them, h j,d represents the hidden features of d dimensions at the jth time step; W j,d Represents the weight vector adjusted during training; function f sco It is implemented through a fully connected layer, whose number of neurons is equal to the number of time steps T;

[0099] Use the softmax function to normalize the scores of a specific dimension so that the sum of the scores at different time steps is 1:

[0100]

[0101] Among them, s j,d 'Indicates the score after normalizing the feature dimension score using the softmax function;

[0102] The scores of all dimensions at the same time step are averaged to obtain the attention weight at each time step. The hidden layer features of the corresponding time step are then multiplied by the attention weight, and the resulting new feature matrix is ​​used as the input of the subsequent layer:

[0103]

[0104] Among them, α j represents the self-attention weight score at the jth time step; h j represents the hidden feature vector at the jth time step; h j wt represents the weighted hidden feature vector at the jth time step;

[0105] The third layer is the second bidirectional long short-term memory network layer, which is responsible for extracting more abstract features from the weighted hidden features transferred above;

[0106] (3) Output layer: The output layer consists of a fully connected layer, which is used to obtain the predicted value of the load component and output the predicted vector of the load component

[0107] Among them, Y (i) represents the prediction result vector of the i-th load component; represents the predicted value of the i-th load component at the start of the prediction time window, t is the target time point for prediction, and h is the prediction range (i.e., the length of the prediction window, e.g., to predict the load for h time steps in the future, the window includes the time interval from th to t);

[0108] It should be noted that the standard empirical mode decomposition algorithm is prone to modal aliasing when decomposing complex signals, resulting in the inclusion of extremely different characteristic time scales in the intrinsic mode function, resulting in poor decomposition effect; the complete adaptive noise integrated empirical mode decomposition algorithm, based on the theoretical basis of the empirical mode decomposition algorithm, effectively overcomes the above shortcomings by adaptively adding Gaussian white noise during the decomposition process, achieving better decomposition effect.

[0109] S4. Superimpose and combine the load component prediction values: Superimpose and combine the prediction vectors of the load components, and reconstruct the vector as the final load forecast result Y based on the improved empirical mode decomposition and bidirectional long short-term memory network prediction model:

[0110] The historical load data is input into the short-term load forecasting model established by combining improved empirical mode decomposition and bidirectional long short-term memory network, and the output is the load forecast value of the day to be forecasted.

[0111] The embodiments of the present invention are described in detail above with reference to the accompanying drawings, but the present invention is not limited thereto. Various changes can be made within the scope of knowledge possessed by those skilled in the art without departing from the spirit of the present invention.

Claims

1. A short-term power load forecasting method combining improved empirical mode decomposition and bidirectional long short-term memory network, characterized in that: The specific implementation steps include the following: S1. Preprocess the historical load data, including missing value filling and outlier correction; S2, using a complete adaptive noise integrated empirical mode decomposition algorithm to decompose the pre-processed historical load data into several intrinsic mode function components; S3. For each intrinsic mode function component, a prediction model based on the self-attention mechanism and bidirectional long short-term memory network is constructed to generate component prediction values; S4. Superimpose all component prediction values ​​to obtain the short-term power load prediction result for the day to be predicted.

2. The short-term power load forecasting method combining improved empirical mode decomposition and bidirectional long short-term memory network according to claim 1 is characterized in that: Missing values ​​are filled using the mean filling method, and the specific formula is: in, represents the average load value on the nth day; x n,i represents the load value at the i-th moment of the n-th day; I is 48, which means 48 pieces of data collected within 24 hours; x defect Indicates the missing values ​​to be filled.

3. The short-term power load forecasting method combining improved empirical mode decomposition and bidirectional long short-term memory network according to claim 2 is characterized in that: Outlier correction includes: Among them, x i represents the mean value at the i-th moment of the day; σ i represents the standard deviation at the i-th moment, and ε represents the set abnormal threshold; If the load value x n,i and The absolute value of the difference is greater than 3σ i ε, it is judged as an abnormal value and corrected. The correction formula is: Among them, x n,i ' represents the correction value at the i-th moment of the n-th day; x m,i Represents the data value at the i-th moment of the same type of day; α and β are the set weight coefficients.

4. The short-term power load forecasting method combining improved empirical mode decomposition and bidirectional long short-term memory network according to claim 3 is characterized in that: The complete adaptive noise ensemble empirical mode decomposition algorithm includes: S21. Add I-order Gaussian white noise to the historical load data y(t): y i (t)=y(t)+ε'N i (t); Among them, N i (t) represents a Gaussian white noise signal that satisfies the standard normal distribution, i = 1, 2, ..., I; ε' represents an adaptive function; y i (t) represents the new data obtained after adding white noise; S22, for the above y i (t) Perform empirical mode decomposition to obtain I first-order modal components, and average them to obtain the first eigenmode function C1'(t): Among them, EMD(y i (t)) represents the new data y obtained after adding white noise i (t) performing an empirical mode decomposition operation, i.e., EMD decomposition; Represents the new data y obtained after adding white noise i (t) The first-order modal component obtained after empirical mode decomposition; r i (t) represents the new data y obtained after adding white noise i (t) is the single decomposition residual obtained after empirical mode decomposition; C1'(t) represents the first intrinsic mode function; S23. For the kth decomposition, first find the remainder r after k-1 decompositions. k-1 (t), and then find the kth eigenmode function according to the following method: Among them, r k-1 (t) represents the total residual after the first k-1 adaptive noise integrated EMD decompositions; C k '(t) represents the intrinsic mode function obtained by the kth decomposition; EMD1(·) represents the first-order mode component obtained by the empirical mode decomposition; EMD k-1 (·) represents the k-1th order modal component obtained by empirical mode decomposition; S24. When the number of extreme points of the residual is ≤ 2, the decomposition is terminated, and K intrinsic mode functions and residual r after K decompositions are obtained. K (t), and based on this we can get the historical load data y(t):

5. The short-term power load forecasting method combining improved empirical mode decomposition and bidirectional long short-term memory network according to claim 4 is characterized in that: The specific conditions for terminating the decomposition are: if the residue r after the kth decomposition k If the number of extreme points of (t) is less than or equal to 2, the decomposition process is stopped.

6. The short-term power load forecasting method combining improved empirical mode decomposition and bidirectional long short-term memory network according to claim 5 is characterized in that: Prediction models based on self-attention mechanism and bidirectional long short-term memory network include: Input layer: The input variable is the load component data mentioned above, the number of components is n, and the load x is extracted using a sliding window to form the input feature vector X: in, is the feature vector extracted from the jth sliding window of the i-th load component; t is the prediction time; h is the prediction horizon; lag is the number of consecutive lagged loads; m is the width of the sliding window, the window sliding step is 1, and the number of sliding windows T = lag + 1-m; Model training layer: It consists of a bidirectional long short-term memory network layer, a self-attention mechanism layer, and a second bidirectional long short-term memory network layer, which is used to extract temporal features and weight key time points; Output layer: The output layer consists of a fully connected layer, which is used to obtain the predicted value of the load component and output the predicted vector of the load component Among them, Y (i) represents the prediction result vector of the i-th load component; It represents the predicted value of the i-th load component at the start time of the prediction time window, t is the target time point of the prediction, and h is the prediction range.

7. The short-term power load forecasting method combining improved empirical mode decomposition and bidirectional long short-term memory network according to claim 6 is characterized in that: The model training layers include: Bidirectional long short-term memory network layer, which extracts hidden features from the load component data. The hidden feature vector can be expressed as: h t =[h t1 ,h t2 ,…,h tm ], t∈[1,T]; Among them, h t is the hidden feature vector; m and T are the dimension of the hidden feature and the number of time steps respectively; The attention layer passes the extracted hidden features to the self-attention mechanism layer, which weights the hidden features through the self-attention mechanism. The process of weighting the hidden features includes: Calculate the score of hidden features of each dimension at different time steps: s j,d =f sco (W j,d [h 1,d ,h 2,d ,…,h T,d ]),d=1,…,m,j=1,…,T; Among them, h j,d represents the hidden features of d dimensions at the jth time step; W j,d Represents the weight vector adjusted during training; function f sco It is implemented through a fully connected layer, whose number of neurons is equal to the number of time steps T; Use the softmax function to normalize the scores of a specific dimension so that the sum of the scores at different time steps is 1: Among them, s j,d 'Indicates the score after normalizing the feature dimension score using the softmax function; The scores of all dimensions at the same time step are averaged to obtain the attention weight at each time step. The hidden layer features of the corresponding time step are then multiplied by the attention weight, and the resulting new feature matrix is ​​used as the input of the subsequent layer: Among them, α j represents the self-attention weight score at the jth time step; h j represents the hidden feature vector at the jth time step; h j wt represents the weighted hidden feature vector at the jth time step; The second bidirectional long short-term memory network layer is responsible for extracting more abstract features from the weighted hidden features passed above.

8. The short-term power load forecasting method combining improved empirical mode decomposition and bidirectional long short-term memory network according to claim 1 is characterized in that: The short-term power load forecast results for the forecast day are: Among them, Y is the short-term power load forecast result of the forecast day; Y (i) is the prediction result vector of the i-th load component.

Images