BiLSTM short-term load prediction method based on ESMD decomposition and error correction strategy optimization

Abstract

The invention belongs to the field of short-term scale power system load data prediction, and discloses a BiLSTM short-term load prediction method based on ESMD decomposition and error correction strategy optimization, and the method comprises the steps: carrying out the adaptive decomposition of load data into a plurality of components through an ESMD decomposition algorithm, and reducing the noise of original data; then predicting the components by using a BiLSTM model to obtain a preliminary prediction result; an error correction strategy ECS calculates an error sequence between the load original data and the preliminary prediction result, and predicts the error sequence by using a BiLSTM model; and adding the error sequence prediction result and the preliminary prediction result to obtain a final prediction result. Meanwhile, according to the method, through an ablation experiment, it is verified that the performance indexes of the proposed method are remarkably improved, and an efficient and accurate solution is provided for short-term scale power system load prediction.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of short-term scale power system load data prediction, in particular to a BiLSTM short-term load prediction method based on ESMD decomposition and error correction strategy optimization. BACKGROUND

[0002] Short-term load forecasting is generally to predict the load demand in the future one hour to one week, which is an important part of power load forecasting. Enhancing the accuracy and robustness of short-term load forecasting has a positive effect on optimizing power generation planning, ensuring the safety and stability of power grids, and accommodating renewable energy. Traditional prediction models mainly include autoregressive moving average model, autoregressive integrated moving average model, multivariate linear regression, Kalman filter and grey prediction model. These models are mainly based on statistics and linear regression analysis, and have high computational efficiency and strong interpretability. However, with the introduction of renewable energy and market mechanisms, load forecasting has gradually developed from traditional linear problems to high-dimensional nonlinear problems, making it difficult for traditional models to capture more complex dynamic characteristics and adapt to large-scale data and computing power requirements. With the development of computer science, deep learning algorithms have developed rapidly, and they have a significant advantage in dealing with highly complex nonlinear relationships between input and output. Common deep learning methods include recurrent neural network (RNN), long short-term memory network (LSTM), gated recurrent unit (GRU) and bidirectional long short-term memory network (BiLSTM). In the field of short-term scale load forecasting, bidirectional long short-term memory network (BiLSTM) has been widely used to capture long-term dependencies and dynamic time-varying characteristics in time series data. However, deep learning models extract data features through multiple layers of nonlinear transformation, and the training process is essentially learning the statistical distribution of input data, so they have high requirements for input data. When the training data contains noise, the model will incorrectly associate noise features with target output, distort the true distribution of the feature space, cause feature extraction bias, and make the model converge to a suboptimal solution. Therefore, the extreme point symmetric mode decomposition algorithm (ESMD) and error correction strategy (ECS) are introduced to reduce noise interference in the original sequence and improve prediction accuracy. In the field of short-term scale power system load forecasting, a more efficient and robust prediction method needs to be built. SUMMARY

[0003] In view of the above defects or improvement needs of the prior art, the present application provides a BiLSTM short-term load prediction method based on ESMD decomposition and error correction strategy optimization (ESMD-BiLSTM-ECS), which significantly improves the performance indicators such as mean square error, mean absolute error, mean absolute percentage error and determination coefficient, providing an efficient and accurate solution for short-term scale power system load forecasting.

[0004] To achieve the above object, according to one aspect of the present application, a BiLSTM short-term load forecasting method based on ESMD decomposition and error correction strategy optimization is provided, which is applied to the field of load data forecasting; comprising the following steps: S1: the original load data value is adaptively decomposed by the ESMD decomposition algorithm to obtain a set of time series with single frequency and residual term sequence; S2: the time series set and residual term sequence obtained in step S1 are divided into 80% data set as training set sequence and 20% data set as test set sequence, and the sequence data is normalized, and the normalized data is divided into sliding window; S3: the time series set and residual term sequence obtained in step S2 are predicted by using BiLSTM prediction module to obtain preliminary load prediction value; S4: combining error correction strategy, calculating the error sequence between the original load data value and the preliminary load prediction value obtained in step S3, and continuing to predict the error sequence by using BiLSTM model to obtain error sequence prediction value; S5: adding the error sequence prediction value obtained in step S4 to the preliminary load prediction value obtained in step S3 to obtain the final load prediction value; S6: comparing the final load prediction value obtained in step S5 with the actual load value, and evaluating the load prediction model.

[0005] As preferred, step S1 specifically comprises the following steps: S11 constructing extreme point extraction and midpoint: for the original load data value extracting all maximum and minimum points, connecting adjacent extreme points to form line segments, calculating the midpoint of each line segment , and supplementing the boundary midpoint at both ends of the signal and ; S12 generating symmetric mean curve: through the to midpoints obtained in step S11, constructing cubic spline interpolation curve, and calculating its mean curve , the formula is: In the formula, is the cubic spline interpolation curve, is the total number of interpolation curves; S13 screening modal component: on the basis of step S12, comparing the mean curve value with the preset error threshold , that is: if ,but For the first time series component Otherwise, the residual sequence Repeat steps S11-S12 as a new signal until the screening conditions are met; Residual sequence The formula is:

[0006] In the formula, This is the original load data; S14 Decompose the residual signal: Calculate the residual term and the residual term Repeat step S13 to extract time series components sequentially. , , Until the remaining signal To obtain a time series set with a single frequency, where there is monotonicity or insufficient poles. IMF i and residual sequence R .

[0007] Preferably, step S15 is also included to determine the optimal number of screenings: setting a range for the number of screenings. By minimizing the variance ratio With the number of screenings Curve determines the optimal number of times , The variance of the residual signal. This represents the variance of the original signal.

[0008] Preferably, in step S1 at time t, all time series components obtained from the decomposition are... With residual sequence R The sum equals the original load value at that moment. The final decomposition result is shown in the following formula: In the formula, This represents the total number of time series components.

[0009] Preferably, the normalized data in step S2 is located in the interval [0,1]. The sliding window size is set to 168, meaning that each input sample includes data from the first 168 time steps, and the output is the target value at the 169th time step.

[0010] Preferably, step S3 specifically includes the following steps: S31: A BiLSTM model is constructed using a two-layer structure to capture the forward and reverse temporal dependencies of sequence data; S32: Output each time series component from step S2 IMF i and residual sequence R Independent BiLSTM models are trained separately, and data is processed in batches during training to calculate loss and backpropagate to update parameters. S33: After each maximum training epoch, the BiLSTM model trained in step S32 enters the evaluation mode. The model parameters corresponding to the maximum training epoch with the smallest mean square error on the test set are selected as the final prediction model, and the predicted values ​​of each time series component and residual sequence of the training set and test set are output. S34: After inverse normalizing the predicted values ​​of each time series component and the residual series obtained in step S33, add them together to obtain the preliminary load forecast value.

[0011] As a preferred embodiment, the BiLSTM model constructed in step S31 has 128 hidden units in the first layer and 64 in the second layer; each BiLSTM layer is followed by a ReLU activation function, and the output layer is a fully connected layer with an output dimension of 1.

[0012] Preferably, during the BiLSTM model training process in step S31, the Adam optimizer is used to update the parameters, the mean squared error is used as the loss function, the learning rate is set to 0.001, and the maximum number of training epochs is set to 100.

[0013] Preferably, step S4 specifically includes the following steps: S41: Subtract the original load data value from the preliminary load forecast value obtained in step S3 to obtain the error sequence; S42: Input the error sequence obtained in step S41 into the BiLSTM model, and repeat steps S31-S34 to obtain the predicted value of the error sequence.

[0014] Preferably, the indicators used to evaluate the load forecasting model in step S6 include mean absolute error, root mean square error, mean absolute percentage error, and coefficient of determination.

[0015] In summary, the technical solutions conceived by this invention have the following beneficial effects compared with the prior art: The method of this invention first adaptively decomposes the load data into several components using the ESMD decomposition algorithm to reduce noise in the original data; then, it uses a BiLSTM model to predict these components to obtain preliminary prediction results; the error correction strategy (ECS) calculates the error sequence between the original load data and the preliminary prediction results, and uses the BiLSTM model to predict the error sequence; the error sequence prediction results are added to the preliminary prediction results to obtain the final prediction results.

[0016] The method of this invention was compared with the BiLSTM model, the BiLSTM model based on ESMD decomposition (ESMD-BiLSTM), and the BiLSTM model based on error correction strategy optimization (BiLSTM-EC) through ablation experiments. The results showed that the proposed method significantly improved various performance indicators (such as mean square error, mean absolute error, mean absolute percentage error, and coefficient of determination), providing an efficient and accurate solution for short-term power system load forecasting. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the ESMD-BiLSTM-ECS model provided by the present invention.

[0018] Figure 2 This is a diagram showing the ESMD decomposition results provided by the present invention.

[0019] Figure 3 This is a typical 72-hour load forecast curve of the BiLSTM model, ESMD-BiLSTM model, BiLSTM-EC model and ESMD-BiLSTM-ECS model in this invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0021] Please see Figure 1 and Figure 2 This invention provides a BiLSTM short-term load forecasting method based on ESMD decomposition and error correction strategy optimization, comprising the following steps: S1: Adaptively decompose the original load dataset using the ESMD decomposition algorithm to obtain a set of time series with a single frequency and a residual term sequence; Step S1 specifically includes the following steps: S11 Extremum Point Extraction and Midpoint Analysis: Extracting and analyzing raw load data values Extract all maximum and minimum points, connect adjacent extreme points to form line segments, and calculate the midpoint of each line segment. And add the midpoint of the boundary at both ends of the signal. and ; S12 generates a symmetrical mean curve: obtained through step S11. arrive The midpoint is constructed using cubic spline interpolation. Draw interpolation curves and calculate their mean curves. Its formula is: (4.1) In the formula, For the first Interpolation curves, This represents the total number of interpolation curves. S13 Filtering Modal Components: Based on step S12, compare the mean curve values. Compared with the preset error threshold ,Right now: like ,but For the first time series component Otherwise, the residual sequence Repeat steps S11-S12 as a new signal until the screening conditions are met; Residual sequence The formula is:

[0022] In the formula, This is the original load data; S14 Decompose the residual signal: Calculate the residual term and the residual term Repeat step S13 to extract time series components sequentially. , , Until the remaining signal To obtain a time series set with a single frequency, where there is monotonicity or insufficient poles. IMF i and residual sequence R .

[0023] S15 Determine the optimal number of filters: Set the range of filters. By minimizing the variance ratio With the number of screenings Curve determines the optimal number of times , The variance of the residual signal. The variance of the original signal; at time t, all time series components obtained from the decomposition With residual sequence R The sum equals the original load value at that moment. The final decomposition result is as follows: In the formula, This represents the total number of time series components.

[0024] S2: Take the time series set and residual term sequence obtained from step S1, use 80% of the dataset as the training set sequence and 20% of the dataset as the test set sequence, and normalize the sequence data. Then divide the normalized data into sliding windows. Specifically, in step S2, the normalized data is located in the interval [0,1]; the sliding window size is set to 168, that is, each input sample includes the data of the first 168 time steps, and the output is the target value of the 169th time step.

[0025] S3: Use the BiLSTM prediction module to predict the time series set and residual term sequence obtained in step S2 to obtain the preliminary load forecast value; Step S3 specifically includes the following steps: S31: A BiLSTM model is constructed using a two-layer structure to capture the forward and reverse temporal dependencies of sequence data; Specifically, the BiLSTM model constructed in step S31 has 128 hidden units in the first layer and 64 in the second layer; each BiLSTM layer is followed by a ReLU activation function, the output layer is a fully connected layer, and the output dimension is 1; during the training of the BiLSTM model in step S31, the Adam optimizer is used to update the parameters, the mean squared error is used as the loss function, the learning rate is set to 0.001, and the maximum training epochs are set to 100. S32: Output each time series component from step S2 IMF i and residual sequence R Independent BiLSTM models are trained separately, and data is processed in batches during training to calculate loss and backpropagate to update parameters. S33: After each maximum training epoch, the BiLSTM model trained in step S32 enters the evaluation mode. The model parameters corresponding to the maximum training epoch with the smallest mean square error on the test set are selected as the final prediction model, and the predicted values ​​of each time series component and residual sequence of the training set and test set are output. S34: After inverse normalizing the predicted values ​​of each time series component and the residual series obtained in step S33, add them together to obtain the preliminary load forecast value.

[0026] S4: Combining the error correction strategy, calculate the error sequence between the original load data value and the preliminary load forecast value obtained in step S3, and continue to use the BiLSTM model to predict the error sequence to obtain the error sequence prediction value. Step S4 specifically includes the following steps: S41: Subtract the original load data value from the preliminary load forecast value obtained in step S3 to obtain the error sequence; S42: Input the error sequence obtained in step S41 into the BiLSTM model, and repeat steps S31-S34 to obtain the predicted value of the error sequence.

[0027] S5: Add the error sequence prediction value obtained in step S3 to the preliminary load prediction value obtained in step S2 to obtain the final load prediction value; S6: Compare the final load prediction result obtained from simulation calculation step S4 with the actual load value to evaluate the load prediction model.

[0028] The metrics used to evaluate the load forecasting model in step S6 include mean absolute error, root mean square error, mean absolute percentage error, and coefficient of determination.

[0029] In step S1, the ESMD decomposition algorithm is Extreme-Point Symmetric Mode Decomposition (ESMD), which is an improved adaptive signal decomposition method. EMD decomposition algorithms suffer from mode aliasing and endpoint effects when dealing with nonlinear and non-stationary signals, and rely on the symmetry of the extreme point envelope, resulting in low computational efficiency. The ESMD decomposition algorithm reduces envelope fitting errors and suppresses endpoint effects through extreme-point symmetric interpolation and a dynamic selection stopping criterion. Simultaneously, it dynamically adjusts the number of selections to avoid over-decomposition, thereby optimizing the accuracy and robustness of mode decomposition.

[0030] Traditional recurrent neural networks (RNNs) suffer from vanishing or exploding gradients when processing long sequences, making it difficult to learn long-term dependencies. Bidirectional long short-term memory (LSTM) networks effectively alleviate this problem by introducing gating mechanisms and memory units to selectively retain or forget information. An LSTM unit consists of two hidden states and three control gates. The two hidden states are... and The LSTM unit is divided into three gates: a short-term gate and a long-term memory gate. The three gates are the forget gate, the input gate, and the output gate. The forget gate controls the degree to which information from the previous time step is retained. Its input includes the hidden state from the previous time step and the input data from the current time step. After activation by the sigmoid function, it outputs a weight value between 0 and 1 to regulate the update of the memory unit. The input gate is also regulated by the sigmoid function and is responsible for filtering new information to be stored in the memory unit. The output gate determines which information from the memory unit will be output to the next time step. These three gating mechanisms work together to regulate the memory update and information output process of the LSTM unit. The formula for calculating the output of the LSTM is as follows:

[0031]

[0032] (4.3)

[0033]

[0034]

[0035] In the formula, Represents the Gate of Oblivion; Represents the input gate; Represents the sigmoid activation function; Indicates the current state of the unit; Indicates a hidden state; Indicates the model input; It is time The output vector at that time; and These represent the gate weight matrix and bias, respectively; and Having a similar form, both doors are made of and Sure; Represents the output gate; This is the final output of the LSTM module.

[0036] Bidirectional Long Short-Term Memory (BiLSTM) networks represent a significant improvement over traditional unidirectional LSTMs. By simultaneously integrating forward and backward LSTM modules, they can more effectively capture contextual information in sequence data, thus significantly enhancing the model's contextual modeling capabilities. Thanks to this bidirectional information processing mechanism, BiLSTM exhibits superior performance in tasks relying on global contextual understanding, such as natural language processing, sequence labeling, and time series analysis. Furthermore, by fusing the output features of both forward and backward LSTMs, BiLSTM can construct more representative feature vectors, an advantage that makes its performance particularly outstanding in numerous practical applications.

[0037] In step S4, the Error Correction Strategy (ECS) refers to the method of detecting and correcting errors in the model output in tasks such as Natural Language Processing (NLP), speech recognition, and machine translation. Its core objective is to improve the robustness and accuracy of the system, ensuring that the final output conforms to grammatical, semantic, or contextual logic. In the field of time series prediction, the error correction strategy uses the deviation between the actual value and the initial predicted value as an error sequence. Then, it predicts based on this error sequence, and superimposes the obtained error sequence prediction value with the initial prediction value to obtain the final prediction result. The calculation formula for the error correction strategy is as follows: (4.4) (4.5) In the formula, Represents the error value; Represents the true value; Represents the initial predicted value; This represents the predicted error value; This represents the final load forecast.

[0038] In summary, the process of predicting short-term power system load using the ESMD-BiLSTM-ECS model of this invention is as follows: Step 1: Data Loading and EMSD Decomposition The original load data values ​​are loaded, and the ESMD decomposition algorithm is used to decompose the original load data to obtain a set of time series with a single frequency. IMF i and residual sequence R ; Step 2: Data Preprocessing and Sliding Window Division The time series set obtained from the decomposition in Step 1 { IMF i and residual sequence RThe dataset was used as a training set sequence with 80% of the dataset and a test set sequence with 20% of the dataset. To improve the generalization ability and stability of the model, the sequence data was normalized, and the normalized data was located in the interval [0,1]. The normalized data was divided into sliding windows with a window size of 168. Each input sample included the data from the first 168 time steps, and the output was the target value at the 169th time step. Step 3: BiLSTM model prediction A BiLSTM model with a two-layer structure is constructed. The first layer has 128 hidden units, and the second layer has 64. This bidirectional structure captures the forward and backward temporal dependencies of the sequence data. Each BiLSTM layer is followed by a ReLU activation function, and the output layer is a fully connected layer with a dimension of 1. During training, the Adam optimizer is used to update parameters, with mean squared error (MSE) as the loss function, a learning rate of 0.001, and a maximum training epoch of 100. The sliding window partitioned in Step 2 is input into the BiLSTM model, and for each... IMF Components and residuals R Independent BiLSTM models are trained separately. During training, the model processes data batch by batch, calculates the loss, and backpropagates to update the parameters. After each epoch, the model enters evaluation mode, selecting the model parameters corresponding to the epoch with the smallest MSE on the test set as the final prediction model, and outputs the prediction results for both the training and test sets. The prediction results of each component are inversely normalized and then summed to obtain the preliminary load prediction value. Step 4: The error correction strategy yields the final prediction result. The error sequence is obtained by subtracting the actual value from the predicted value. A BiLSTM model is then used to predict the error sequence, with the same sliding window settings as in Step 2 and the same hyperparameter settings as in Step 3. The model is trained to obtain the predicted error sequence values. Finally, the predicted error sequence values ​​are added to the initial load forecast values ​​to obtain the final load forecast values.

[0039] Step 5: Ablation experiment design.

[0040] The proposed model was compared to the one proposed by progressively removing modules from the ESMD-BiLSTM-ECS model through ablation experiments. To verify the effectiveness of the ESMD algorithm, a control group was designed, consisting of an ESMD-BiLSTM-ECS model, a BiLSTM-ECS model, and a BiLSTM-ECS model. Similarly, to verify the effectiveness of the error correction strategy (ECS), a control group was designed, consisting of an ESMD-BiLSTM-ECS model, an ESMD-BiLSTM model, and a BiLSTM-ECS model.

[0041] Step 6: Model Evaluation; To comprehensively evaluate the predictive performance and computational efficiency of the experimental design model, this invention selected four prediction evaluation metrics: Mean Absolute Error (MAE), Root Mean Square Error (RMSE), Mean Absolute Percentage Error (MAPE), and Coefficient of Determination (R²). 2 ).

[0042] (4.6) (4.7) (4.8) (4.9) In the formula, , These represent the observed value and the predicted value, respectively. This represents the mean of the observed values; This is the length of the test set.

[0043] Specifically, this invention uses hourly-scale power system load data from Yunnan Province as an example to verify the effectiveness of the invention. The load forecasting method of this invention and... Figure 1 The overall process of the ESMD-BiLSTM-ECS model shown is used for load forecasting to demonstrate the advantages of this invention.

[0044] The example uses hourly data to obtain hourly power system load data for Yunnan Province from July 1, 2020 to June 30, 2021.

[0045] First, the power system load data is adaptively decomposed using the ESMD algorithm to reduce noise in the data, resulting in nine time series with single frequencies. IMF 1 ~ IMF 9 and residual sequence R The decomposition results are as follows Figure 2 As shown.

[0046] The decomposed sequences are normalized to fall within the [0, 1] interval to eliminate the influence of different units. The dataset is divided into a training set and a test set, with the training set comprising 80% of the data and the remaining 20% ​​as the test set. A sliding window method is used to partition the time series data, converting it into a format suitable for training the prediction model. The sliding window size is set to 168, with each input sample including data from the first 168 time steps, and the output being the target value at the 169th time step.

[0047] A BiLSTM model with a two-layer structure is constructed. The first layer has 128 hidden units, and the second layer has 64. Each BiLSTM layer is followed by a ReLU activation function, and the output layer is a fully connected layer with a dimension of 1. During training, the Adam optimizer is used to update the parameters, with mean squared error (MSE) as the loss function, a learning rate of 0.001, and a maximum training epoch of 100. The divided sliding window is input into the BiLSTM model, and for each... IMF Components and residuals R Independent BiLSTM models are trained separately. The model parameters corresponding to the epoch with the smallest MSE on the test set are selected as the final prediction model, and the prediction results for both the training and test sets are output. The prediction results of each component are inversely normalized and then summed to obtain the preliminary load prediction value.

[0048] The difference between the original load sequence value and the preliminary forecast result is calculated to obtain the error sequence. The error sequence is then predicted using a BiLSTM model with the same hyperparameter settings to obtain the error sequence prediction value. Finally, the error sequence prediction value is added to the preliminary load forecast value to obtain the final load forecast value.

[0049] To evaluate the effectiveness of the prediction model proposed in this invention, an ablation experiment was conducted. BiLSTM, ESMD-BiLSTM, and BiLSTM-EC models were used as comparative models, while ESMD-BiLSTM-ECS was the model proposed in this invention. Three-day (72-hour) load prediction results, including extreme points, were selected. The prediction results are as follows: Figure 3 As shown, the proposed model fits the load curve well and exhibits excellent performance in extreme value prediction. Mean absolute error (MAE), mean squared error (RMSE), mean absolute percentage error (MAPE), and coefficient of determination (R²) are used to measure the performance. 2 The predictive performance of the model was evaluated by the indicators. The calculation results of the ESMD-BiLSTM-ECS model and other comparative models are shown in Table 1.

[0050]

[0051] The results in the table show that the proposed ESMD-BiLSTM-ECS model achieves optimal results in all four evaluation metrics, validating the effectiveness and superiority of the proposed model in short-term power system load forecasting. Specifically, compared with the BiLSTM model, the ESMD-BiLSTM model achieves better results in RMSE, MAE, MAPE, and R... 2 The ESMD-BiLSTM-ECS model improved by 34.69%, 40.46%, 39.70%, and 1.41% respectively compared to the BiLSTM-EC model in terms of RMSE, MAE, MAPE, and R. 2 The results showed improvements of 7.22%, 8.38%, 6.46%, and 0.16% respectively, demonstrating that ESMD decomposition of the data sequence reduces noise interference and helps improve load forecasting accuracy. The BiLSTM-EC model, compared to the BiLSTM model, also improved RMSE, MAE, MAPE, and R... 2 The ESMD-BiLSTM-ECS model improved by 28.84%, 35.18%, 35.64%, and 1.26% respectively compared to the ESMD-BiLSTM model in terms of RMSE, MAE, MAPE, and R. 2 The accuracy rates were improved by 0.35%, 0.26%, 0.17%, and 0.01% respectively, demonstrating that the error correction strategy can further uncover the connections between data and improve prediction accuracy.

[0052] Overall, the BiLSTM short-term load forecasting method based on ESMD decomposition and error correction strategy optimization has excellent performance and robustness in power system load forecasting in Yunnan Province, providing an efficient and accurate solution for short-term power system load forecasting.

[0053] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A BiLSTM short-term load forecasting method based on ESMD decomposition and error correction strategy optimization, characterized in that, This method is applied to the field of load data forecasting; it includes the following steps: S1: Adaptively decompose the original load data values ​​using the ESMD decomposition algorithm to obtain a set of time series with a single frequency and a residual term sequence; S2: Take the time series set and residual term sequence obtained from step S1, use 80% of the dataset as the training set sequence and 20% of the dataset as the test set sequence, and normalize the sequence data. Then divide the normalized data into sliding windows. S3: Use the BiLSTM prediction module to predict the time series set and residual term sequence obtained in step S2 to obtain the preliminary load forecast value; S4: Combining the error correction strategy, calculate the error sequence between the original load data value and the preliminary load forecast value obtained in step S3, and continue to use the BiLSTM model to predict the error sequence to obtain the error sequence prediction value. S5: Add the error sequence prediction value obtained in step S4 to the preliminary load prediction value obtained in step S3 to obtain the final load prediction value; S6: Compare the final load forecast value obtained from simulation calculation step S5 with the actual load value to evaluate the load forecast model.

2. The BiLSTM short-term load forecasting method based on ESMD decomposition and error correction strategy optimization as described in claim 1, characterized in that, Step S1 specifically includes the following steps: S11 Extremum Point Extraction and Midpoint Analysis: Extracting and analyzing raw load data values Extract all maximum and minimum points, connect adjacent extreme points to form line segments, and calculate the midpoint of each line segment. And add the midpoint of the boundary at both ends of the signal. and ; S12 generates a symmetrical mean curve: obtained through step S11. arrive The midpoint is constructed using cubic spline interpolation. Draw interpolation curves and calculate their mean curves. Its formula is: In the formula, For the first Interpolation curves, This represents the total number of interpolation curves. S13 Filtering Modal Components: Based on step S12, compare the mean curve values. Compared with the preset error threshold ,Right now: like ,but For the first time series component Otherwise, the residual sequence Repeat steps S11-S12 as a new signal until the screening conditions are met; Residual sequence The formula is: In the formula, This is the original load data; S14 Decompose the residual signal: Calculate the residual term and the residual term Repeat step S13 to extract time series components sequentially. , , Until the remaining signal To obtain a time series set with a single frequency, where there is monotonicity or insufficient poles. IMF i and residual sequence R .

3. The BiLSTM short-term load forecasting method based on ESMD decomposition and error correction strategy optimization as described in claim 2, characterized in that, It also includes step S15, which determines the optimal number of filters: setting a range for the number of filters. By minimizing the variance ratio Number of screenings Curve determines the optimal number of times , The variance of the residual signal. This represents the variance of the original signal.

4. The BiLSTM short-term load forecasting method based on ESMD decomposition and error correction strategy optimization as described in claim 2, characterized in that, In step S1, at time t, all time series components obtained from the decomposition With residual sequence R The sum equals the original load value at that moment. The final decomposition result is shown in the following formula: In the formula, This represents the total number of time series components.

5. A BiLSTM short-term load forecasting method based on ESMD decomposition and error correction strategy optimization as described in claim 2, 3, or 4, characterized in that, The normalized data in step S2 is located in the interval [0,1]. The sliding window size is set to 168, meaning that each input sample includes data from the first 168 time steps, and the output is the target value at the 169th time step.

6. The BiLSTM short-term load forecasting method based on ESMD decomposition and error correction strategy optimization as described in claim 5, characterized in that, Step S3 specifically includes the following steps: S31: A BiLSTM model is constructed using a two-layer structure to capture the forward and reverse temporal dependencies of sequence data; S32: Output each time series component from step S2 IMF i and residual sequence R Independent BiLSTM models are trained separately, and data is processed in batches during training to calculate loss and backpropagate to update parameters. S33: After each maximum training epoch, the BiLSTM model trained in step S32 enters the evaluation mode. The model parameters corresponding to the maximum training epoch with the smallest mean square error on the test set are selected as the final prediction model, and the predicted values ​​of each time series component and residual sequence of the training set and test set are output. S34: After inverse normalizing the predicted values ​​of each time series component and the residual series obtained in step S33, add them together to obtain the preliminary load forecast value.

7. The BiLSTM short-term load forecasting method based on ESMD decomposition and error correction strategy optimization as described in claim 5, characterized in that, The BiLSTM model constructed in step S31 has 128 hidden units in the first layer and 64 in the second layer; each BiLSTM layer is followed by a ReLU activation function, and the output layer is a fully connected layer with an output dimension of 1.

8. The BiLSTM short-term load forecasting method based on ESMD decomposition and error correction strategy optimization as described in claim 5, characterized in that, During the BiLSTM model training process in step S31, the Adam optimizer is used to update the parameters, the mean squared error is used as the loss function, the learning rate is set to 0.001, and the maximum number of training epochs is set to 100.

9. A BiLSTM short-term load forecasting method based on ESMD decomposition and error correction strategy optimization as described in claim 6, 7, or 8, characterized in that, Step S4 specifically includes the following steps: S41: Subtract the original load data value from the preliminary load forecast value obtained in step S3 to obtain the error sequence; S42: Input the error sequence obtained in step S41 into the BiLSTM model, and repeat steps S31-S34 to obtain the predicted value of the error sequence.

10. The BiLSTM short-term load forecasting method based on ESMD decomposition and error correction strategy optimization as described in claim 1, characterized in that, The metrics used to evaluate the load forecasting model in step S6 include mean absolute error, root mean square error, mean absolute percentage error, and coefficient of determination.

Images