Residential area charging pile load prediction method based on deep learning and load peak correction

Abstract

The invention relates to a residential area charging pile load prediction method based on deep learning and load peak correction, and the method comprises the following steps: A), carrying out the preprocessing of residential area charging data, including abnormal value identification, missing value completion, standardization and the like; b) adopting MIC (Microphone) to screen characteristics with relatively high load correlation, and taking the characteristics as input of a load peak value and load trend prediction model; c) constructing an LSTM load peak prediction model; d) carrying out CEEMDAN preliminary decomposition on the load trend, carrying out aggregation reconstruction on each sequence by using K-Means, and then carrying out secondary decomposition on high frequency by using VMD; e) a sequence obtained through secondary reconstruction is transmitted into an Informer prediction model; f) replacing a load peak value to complete curve shape correction; and G) calculating an error index, and evaluating the performance of the model. The method has the advantages that charging behaviors, historical loads, external characteristics and the like are selected as input according to the charging load characteristics of the residential area, rules in load data are fully mined through secondary decomposition and reconstruction, the load peak value is corrected, the prediction precision is improved, and therefore the transformer capacity of the power distribution network is optimized.

Description

Technical Field

[0001] The present invention relates to the technical field of residential distribution network and charging pile load prediction, and in particular to a method for predicting charging pile load in residential areas based on deep learning and load peak correction. Background Art

[0002] With the popularity of electric vehicles and new energy vehicles, a large number of private charging piles have been added to the power grid, posing major challenges to the planning, scheduling and control of the power system in residential areas. With the continuous increase in EV penetration, the access of a large number of charging piles may lead to problems such as increased peak-to-valley differences in the daily load curve of the distribution network and decreased node voltage, which will cause a greater impact on the power grid. Therefore, by analyzing historical and unopened load data and conducting residential charging load forecasting, it can provide important support for the vehicle-grid interactive scheduling and operation strategy, guide the planning and construction of the power grid, optimize the power configuration, and thus formulate a scientific operation and maintenance strategy.

[0003] As the basic support for grid dispatching, customer transformation and relationship planning, and vehicle-grid interactive transactions, load forecasting plays a decisive role in the new power system. At present, the load forecasting of charging piles in residential areas is mainly divided into model-driven and data-driven methods. Model-driven methods include Markov chain, Monte Carlo, etc., which consider the parameters of EV, including state of charge (SOC), driving distance, power consumption and charging duration, and infer the charging behavior of car owners through mathematical models and the spatiotemporal distribution of charging loads. Since the model-driven method lacks real charging load data, it mostly uses mathematical modeling or simulation to simulate real behavior, so the authenticity of its prediction results is limited. Data-driven research methods include statistics, machine learning, and deep learning, which use historical charging load data to predict the load in a specific area, which can better reflect the authenticity and reliability of charging load prediction results.

[0004] Most of the above studies are aimed at predicting the charging load of public charging stations, or simulated predictions based on probability models and simulated occurrence and charging processes, while for private charging stations, there are few studies that directly use real meter data to predict the charging load of residential clusters. Due to the early construction of old communities, the line and node capacity cannot meet the load growth caused by the access of a large number of EV charging loads to the distribution network. In the early stages of construction of new communities, the transformer capacity must also be considered to meet the surge in load caused by the increase in the number of charging piles. Summary of the invention

[0005] In view of the shortcomings of the prior art, after obtaining the charging pile data of the target residential area, the present invention performs data preprocessing, divides the charging load value into two parts: load peak and load curve trend, uses MIC to screen features with a large correlation with the charging load, and inputs them into the LSTM model to obtain the prediction result of the charging load peak of the target community. After the load trend is initially decomposed by CEEMDAN, each sequence is aggregated and reconstructed using K-Means, and then the high frequency is secondary decomposed using VMD. The reconstructed sequence is input into the Informer model for prediction, and the peak in the prediction result is replaced with the result of the load peak prediction in the previous step, and the curve shape is corrected. Finally, the error index is calculated, the model performance is evaluated, and the trained model is used to predict the charging load of the target residential area.

[0006] The method for predicting charging pile load in residential areas based on deep learning and load peak correction comprises the following steps:

[0007] Step 1: Data preprocessing: Due to various reasons such as network and electricity meter failure, the original charging data may be missing or abnormal. In order to improve the prediction accuracy and speed up the training, the data needs to be preprocessed, including missing value filling, outlier identification and replacement, and normalization.

[0008] Step 2: Feature selection: Select the influencing features of the daily peak load data of residential charging, calculate the correlation between different influencing factors and the daily peak load through MIC, and select the peak load features. For load trend data, take 15 minutes as the sampling point, and select the behavioral features with greater correlation among the factors affecting user charging behavior;

[0009] Step 3: Train the LSTM load peak prediction model: Build an LSTM deep learning model based on the feature indicators selected in step 2 to learn the time series information in the daily charging load peak data;

[0010] Step 4: Secondary decomposition and reconstruction of load trend curve: CEEMDAN is used to perform the initial adaptive decomposition of charging load data to obtain multiple subsequences, and the sample entropy of each subsequence is calculated. K-Means is used to divide the sequence into low, medium, and high frequency components according to the sample entropy calculation results, and VMD is used to perform secondary decomposition on the high frequency components obtained by clustering, where the number of VMD decompositions is obtained according to the calculated center frequency;

[0011] Step 5: Train the load trend informer model: divide the data set, construct the input of the informer model according to the sliding window method, and predict the charging trend curve;

[0012] Step 6: Load curve shape correction: Replace the load peak point in the charging load curve of the target residential area predicted by the Informer model with the peak value predicted by the LSTM model in step 3;

[0013] Step 7: Model evaluation: By calculating the values ​​of different error evaluation indicators under different prediction scenarios, the prediction accuracy of the model is comprehensively evaluated;

[0014] As a preference, step 1 data preprocessing process:

[0015] (1): Identification and correction of outliers: Combine the charging behavior of charging pile users and the capacity of charging piles to set the maximum load and the longest and lowest charging time thresholds. For charging behavior data that exceeds or is significantly lower than this range, make an abnormal data mark and make corrections. The charging load of charging piles in fixed residential areas has strong periodicity and regularity, and generally does not fluctuate violently. The outliers are corrected by mean filling. The specific formula is as follows:

[0016]

[0017] Where: y(k,t) represents the charging load value of the target residential area at time t on the kth day;

[0018] (2): Missing value filling: In order to maintain the continuity of the time series, the missing values ​​in the corrected data are generally filled based on the charging load of the adjacent charging date or similar days. The specific formula is as follows:

[0019] Y(d,t)=ω1Y(d,t-1)+ω2Y(d,t+1)+ω3Y(d-1,t)+ω4Y(d+1,t) (2)

[0020] Where: Y(d,t) represents the charging load value of the target residential area at time t on the dth day; ω is the weight coefficient of different similar days;

[0021] (3): Standardization processing: In the data standardization process, the maximum and minimum normalization method is selected. The core idea of ​​this method is to find the maximum and minimum values ​​in the data set, scale the data of different dimensions according to a certain ratio, and map them to [0,1]. The advantage of maximum and minimum normalization is that it retains the distribution form of the original data and is suitable for situations where data needs to be mapped to a specific range. In the study, maximum and minimum normalization was selected to ensure the consistency and comparability of the data in subsequent analysis and modeling. The calculation formula for maximum and minimum normalization is:

[0022]

[0023] Where: X' is the normalized charging load; X represents the original load data; Xmax , X min Respectively represent the maximum and minimum values ​​in the load sequence;

[0024] When the model is trained, the predicted value is still in the interval [0,1]. At this time, the predicted value needs to be denormalized to restore it to its original order of magnitude. The denormalization calculation formula is as follows:

[0025] X=X′*(X max -X min )+X min (4)

[0026] Where: X represents the load data after denormalization; X' represents the predicted load value initially obtained by LSTM; X max , X min Represent the maximum and minimum values ​​in the data sequence respectively.

[0027] As a preference, step 2 feature selection process:

[0028] The correlation between different influencing factors and charging load is quantitatively analyzed, and the maximum information coefficient (MIC) is used to screen the features that show strong correlation with charging load to improve the quality of input data in the model. MIC is a correlation analysis method based on mutual information, which can quantify the linear and nonlinear correlation between two variables. It has extremely high universality and fairness. The detailed calculation formula is as follows:

[0029]

[0030] Where: a and b are the interval numbers in the x and y directions respectively; s is a variable whose size is equal to 0.6 times the amount of data; p(x,y) is the joint probability distribution of x and y; p(x) and p(y) are the marginal probability distributions of x and y respectively; the MIC value ranges from 0 to 1, and the larger the value, the stronger the correlation between the two variables;

[0031] In order to accurately predict the changes in the charging pile load in residential areas, it is necessary to consider the factors affecting the charging pile load from multiple dimensions. For the daily charging load peak of the target residential area, MIC is used to calculate the correlation between the number of charging piles installed in the residential area on the corresponding date, the average temperature, air pressure, precipitation, electricity price, season, holidays, and other influencing characteristics and the daily load peak.

[0032] For the charging load trend with 15-minute sampling points, MIC is used to calculate the correlation between the number of charging piles that are charging and starting to charge and the load at each sampling point. By calculating the correlation between each feature and the load, the feature with the strongest correlation is selected as the input variable of the prediction model, which can not only improve the prediction accuracy but also reduce the calculation complexity of the model.

[0033] As a preferred method, step 3 trains the LSTM daily load peak prediction model:

[0034] LSTM selectively adds new information and forgets previously accumulated information through the synergistic effect of three gating mechanisms (including forget gate, input gate and output gate). The LSTM layer can learn the characteristics of multivariate historical data and predict multivariate loads.

[0035] Confirm the input time series data length of the load peak prediction model, that is, the historical load peak, and set the structure of the LSTM prediction model, including the optimal number of layers, number of neurons, optimizer, and activation function type of the model.

[0036] As a preferred method, step 4 is to reconstruct the load trend curve by secondary decomposition:

[0037] (1): First, the charging load sequence of the target residential area is adaptively decomposed using the fully integrated empirical mode decomposition with adaptive noise (CEEMDAN) to obtain multiple subsequences. Then, the sample entropy of each subsequence is calculated respectively, and the complexity of each subsequence is evaluated by the size of the sample entropy value.

[0038] (2): Use the K-Means clustering algorithm to cluster the sample entropy values ​​of each subsequence in (1). By setting the K-Means clustering parameter to 3, the corresponding subsequences are clustered into low, medium, and high frequency components. Separating components of different complexity helps improve the accuracy of prediction.

[0039] (3): Since the high frequency after aggregation is still complex and difficult to predict, the variational mode decomposition (VMD) algorithm is used to perform secondary decomposition of the high frequency components, and the corresponding central frequencies are calculated for different decomposition numbers K. When the central frequency remains stable as the K value increases, the K value is determined to be the optimal decomposition number of the VMD algorithm, and then K subsequences of the secondary decomposition of the high frequency components are obtained, thereby further reducing data complexity and improving prediction accuracy;

[0040] (4): The K subsequences (K+2 subsequences) obtained by secondary decomposition of the low-frequency component and the medium-frequency component obtained by aggregation in (2) and the medium-high-frequency component in (3) are combined and reconstructed as the input for load trend prediction;

[0041] As a preference, step 5 trains the load trend Informer model:

[0042] (1) Charging data set division: Use the sliding window method to sequentially use the data set with time series characteristics as the input of the model, and define the number of data input in a single sliding window as the time step, and the number of prediction results output by a single sliding window as the prediction step. The overall data set will be divided into training set, validation set, and test set in a ratio of 6:2:2. The training set is used to train the parameters of the model, the validation set is used to verify whether the model is reasonable, and the test set is used for actual prediction.

[0043] (2) Constructing input matrices: For the training set, validation set, and test set, two input matrices need to be constructed first, which are the encoder input x en and decoder input x de , used to predict a sequence after a specific time.

[0044] (3) Model training: Set the optimal learning rate, step size, batch size, and number of training iterations for the Informer model. Use the user charging line features extracted in step 2 and the K+2 subsequences obtained after aggregation and reconstruction in step 4 as the input of the prediction model. Construct K+2 Informer models for prediction respectively. Perform weighted superposition of the prediction results of each model through the fully connected layer to obtain the final charging load trend prediction result.

[0045] (4) Adam optimization algorithm: Select the Adam algorithm to optimize the Informer model. During the model training process, the learning rate of the Adam algorithm will be adaptively adjusted according to the model training situation. It is a further improvement on the gradient descent method, making the model more stable during the training process.

[0046] As a preference, step 6 load curve shape correction:

[0047] The charging load trend curve obtained in step 5 can reflect the total load demand of charging piles in the target residential area in the future time period, but the peak prediction is not accurate enough during concentrated charging at night. In order to improve the accuracy of daily load peak prediction and ensure the safety of residential transformers, the load peak in the predicted load curve obtained in step 5 is replaced with the target residential area charging daily load peak obtained in step 3 to obtain the final charging load prediction value of the target residential area.

[0048] As a preference, step 7 model evaluation:

[0049] The root mean square error (RMSE), mean absolute error (MAE), and goodness of fit (R 2) is used to evaluate the prediction performance of the model and the accuracy of the model prediction results. The calculation formulas of each indicator are as follows:

[0050]

[0051] Where:y i is the actual charging load value at the i-th moment, is the corrected charging load prediction value; n is the number of training samples. The smaller the values ​​of MAE and RMSE, the better the R 2 The closer the value is to 1, the higher the accuracy of the model prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 This is a flow chart of the solution proposed by the present invention; Figure 2 Reconstruct the flow chart for CEEMDAN-VMD secondary decomposition; Figure 3 Predict the flow chart for Informer; DETAILED DESCRIPTION

[0053] The technical solutions in the embodiments of the present invention will be clearly and completely described below in combination with the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0054] Example 1: A method for predicting charging pile load in residential areas based on deep learning and load peak correction. The method comprises the following steps:

[0055] Step 1: Data preprocessing: Due to various reasons such as network and electricity meter failures, the original charging data may be missing or abnormal. In order to improve the prediction accuracy and speed up the training, the data needs to be preprocessed, including missing value filling, outlier identification and replacement, and normalization.

[0056] Step 2: Feature selection: Select the influencing features of the daily peak load data of residential charging, calculate the correlation between different influencing factors and the daily peak load through MIC, and select the peak load features. For load trend data, take 15 minutes as the sampling point, and select the behavioral features with greater correlation among the factors affecting user charging behavior;

[0057] Step 3: Train the LSTM load peak prediction model: Build an LSTM deep learning model based on the feature indicators selected in step 2 to learn the time series information in the daily charging load peak data;

[0058] Step 4: Secondary decomposition and reconstruction of load trend curve: CEEMDAN is used to perform the initial adaptive decomposition of charging load data to obtain multiple subsequences, and the sample entropy of each subsequence is calculated. K-Means is used to divide the sequence into low, medium, and high frequency components according to the sample entropy calculation results, and VMD is used to perform secondary decomposition on the high frequency components obtained by clustering, where the number of VMD decompositions is obtained according to the calculated center frequency;

[0059] Step 5: Train the load trend informer model: divide the data set, construct the input of the informer model according to the sliding window method, and predict the charging trend curve;

[0060] Step 6: Load curve shape correction: Replace the load peak point in the charging load curve of the target residential area predicted by the Informer model with the peak value predicted by the LSTM model in step 3;

[0061] Step 7: Model evaluation: By calculating the values ​​of different error evaluation indicators under different prediction scenarios, the prediction accuracy of the model is comprehensively evaluated;

[0062] Specific steps:

[0063] (I) Data preprocessing:

[0064] (1): Identification and correction of outliers: Combine the charging behavior of charging pile users and the capacity of charging piles to set the maximum load and the longest and lowest charging time thresholds. For charging behavior data that exceeds or is significantly lower than this range, make an abnormal data mark and make corrections. The charging load of charging piles in fixed residential areas has strong periodicity and regularity, and generally does not fluctuate violently. The outliers are corrected by mean filling. The specific formula is as follows:

[0065]

[0066] Where: y(k,t) represents the charging load value of the target residential area at time t on the kth day;

[0067] (2): Missing value filling: In order to maintain the continuity of the time series, the missing values ​​in the corrected data are generally filled based on the charging load of the adjacent charging date or similar days. The specific formula is as follows:

[0068] Y(d,t)=ω1Y(d,t-1)+ω2Y(d,t+1)+ω3Y(d-1,t)+ω4Y(d+1,t) (8)

[0069] Where: Y(d,t) represents the charging load value of the target residential area at time t on the dth day; ω is the weight coefficient of different similar days;

[0070] (3): Standardization processing: In the data standardization process, the maximum and minimum normalization method is selected. The core idea of ​​this method is to find the maximum and minimum values ​​in the data set, scale the data of different dimensions according to a certain ratio, and map them to [0,1]. The advantage of maximum and minimum normalization is that it retains the distribution form of the original data and is suitable for situations where data needs to be mapped to a specific range. In the study, maximum and minimum normalization was selected to ensure the consistency and comparability of the data in subsequent analysis and modeling. The calculation formula for maximum and minimum normalization is:

[0071]

[0072] Where: X' is the normalized charging load; X represents the original load data; X max , X min Respectively represent the maximum and minimum values ​​in the load sequence;

[0073] When the model is trained, the predicted value is still in the interval [0,1]. At this time, the predicted value needs to be denormalized to restore it to its original order of magnitude. The denormalization calculation formula is as follows:

[0074] X=X′*(X max -X min )+X min (10)

[0075] Where: X represents the load data after denormalization; X' represents the predicted load value initially obtained by LSTM; X max , X min Represent the maximum and minimum values ​​in the data sequence respectively.

[0076] (II) Feature selection:

[0077] The correlation between different influencing factors and charging load is quantitatively analyzed, and the maximum information coefficient (MIC) is used to screen the features that show strong correlation with charging load to improve the quality of input data in the model. MIC is a correlation analysis method based on mutual information, which can quantify the linear and nonlinear correlation between two variables. It has extremely high universality and fairness. The detailed calculation formula is as follows:

[0078]

[0079] Where: a and b are the interval numbers in the x and y directions respectively; s is a variable whose size is equal to 0.6 times the amount of data; p(x,y) is the joint probability distribution of x and y; p(x) and p(y) are the marginal probability distributions of x and y respectively; the MIC value ranges from 0 to 1, and the larger the value, the stronger the correlation between the two variables;

[0080] In order to accurately predict the changes in the charging pile load in residential areas, it is necessary to consider the factors affecting the charging pile load from multiple dimensions. For the daily charging load peak of the target residential area, MIC is used to calculate the correlation between the number of charging piles installed in the residential area on the corresponding date, the average temperature, air pressure, precipitation, electricity price, season, holidays, and other influencing characteristics and the daily load peak.

[0081] For the charging load trend with 15-minute sampling points, MIC is used to calculate the correlation between the number of charging piles that are charging and starting to charge and the load at each sampling point. By calculating the correlation between each feature and the load, the feature with the strongest correlation is selected as the input variable of the prediction model, which can not only improve the prediction accuracy but also reduce the calculation complexity of the model.

[0082] (III) Training LSTM load peak prediction model:

[0083] LSTM selectively adds new information and forgets previously accumulated information through the synergistic effect of three gating mechanisms (including forget gate, input gate and output gate). The LSTM layer can learn the characteristics of multivariate historical data and predict multivariate loads.

[0084] Confirm the input time series data length of the load peak prediction model, that is, the historical load peak, and set the structure of the LSTM prediction model, including the optimal number of layers, number of neurons, optimizer, and activation function type of the model.

[0085] (IV) Secondary decomposition and reconstruction of load trend curve:

[0086] (1): First, the charging load sequence of the target residential area is adaptively decomposed using the fully integrated empirical mode decomposition with adaptive noise (CEEMDAN) to obtain multiple subsequences. Then, the sample entropy of each subsequence is calculated respectively, and the complexity of each subsequence is evaluated by the size of the sample entropy value.

[0087] (2): Use the K-Means clustering algorithm to cluster the sample entropy values ​​of each subsequence in (1). By setting the K-Means clustering parameter to 3, the corresponding subsequences are clustered into low, medium, and high frequency components. Separating components of different complexity helps improve the accuracy of prediction.

[0088] (3): Since the high frequency after aggregation is still complex and difficult to predict, the variational mode decomposition (VMD) algorithm is used to perform secondary decomposition of the high frequency components, and the corresponding central frequencies are calculated for different decomposition numbers K. When the central frequency remains stable as the K value increases, the K value is determined to be the optimal decomposition number of the VMD algorithm, and then K subsequences of the secondary decomposition of the high frequency components are obtained, thereby further reducing data complexity and improving prediction accuracy;

[0089] (4): The K subsequences (K+2 subsequences) obtained by secondary decomposition of the low-frequency component and the medium-frequency component obtained by aggregation in (2) and the medium-high-frequency component in (3) are combined and reconstructed as the input for load trend prediction;

[0090] (V) Training Load Trend Informer Model:

[0091] (1) Charging data set division: Use the sliding window method to sequentially use the data set with time series characteristics as the input of the model, and define the number of data input in a single sliding window as the time step, and the number of prediction results output by a single sliding window as the prediction step. The overall data set will be divided into training set, validation set, and test set in a ratio of 6:2:2. The training set is used to train the parameters of the model, the validation set is used to verify whether the model is reasonable, and the test set is used for actual prediction.

[0092] (2) Constructing input matrices: For the training set, validation set, and test set, two input matrices need to be constructed first, which are the encoder input x en and decoder input x de , used to predict a sequence after a specific time.

[0093] (3) Model training: Set the optimal learning rate, step size, batch size, and number of training iterations for the Informer model. Use the user charging line features extracted in step 2 and the K+2 subsequences obtained after aggregation and reconstruction in step 4 as the input of the prediction model. Construct K+2 Informer models for prediction respectively. Perform weighted superposition of the prediction results of each model through the fully connected layer to obtain the final charging load trend prediction result.

[0094] (4) Adam optimization algorithm: Select the Adam algorithm to optimize the Informer model. During the model training process, the learning rate of the Adam algorithm will be adaptively adjusted according to the model training situation. It is a further improvement on the gradient descent method, making the model more stable during the training process.

[0095] (VI) Load curve shape correction:

[0096] The charging load trend curve obtained in step 5 can reflect the total load demand of charging piles in the target residential area in the future time period, but the peak prediction is not accurate enough during concentrated charging at night. In order to improve the accuracy of daily load peak prediction and ensure the safety of residential transformers, the load peak in the predicted load curve obtained in step 5 is replaced with the target residential area charging daily load peak obtained in step 3 to obtain the final charging load prediction value of the target residential area.

[0097] 7. Model evaluation

[0098] The root mean square error (RMSE), mean absolute error (MAE), and goodness of fit (R 2 ) is used to evaluate the prediction performance of the model and the accuracy of the model prediction results. The calculation formulas of each indicator are as follows:

[0099]

[0100] Where:y i is the actual charging load value at the i-th moment, is the corrected charging load prediction value; n is the number of training samples. The smaller the values ​​of MAE and RMSE, the better the R 2 The closer the value is to 1, the higher the accuracy of the model prediction.

Claims

1. A method for predicting charging pile load in residential areas based on deep learning and load peak correction, characterized in that: The following steps are involved: Step 1: Data preprocessing: Due to various reasons such as network and electricity meter failure, the original charging data may be missing or abnormal. In order to improve the prediction accuracy and speed up the training, the data needs to be preprocessed, including missing value filling, outlier identification and replacement, and normalization. Step 2: Feature selection: Select the influencing features of the daily peak load data of residential charging, calculate the correlation between different influencing factors and the daily peak load through MIC, select the peak load features, and for the load trend data, take 15 minutes as the sampling point, and select the behavioral features with greater correlation among the factors affecting user charging behavior; Step 3: Train the LSTM load peak prediction model: Build an LSTM deep learning model based on the feature indicators selected in step 2 to learn the time series information in the daily charging load peak data; Step 4: Secondary decomposition and reconstruction of load trend curve: CEEMDAN is used to perform the initial adaptive decomposition of charging load data to obtain multiple subsequences, and the sample entropy of each subsequence is calculated. K-Means is used to divide the sequence into low, medium, and high frequency components according to the sample entropy calculation results, and VMD is used to perform secondary decomposition on the high frequency components obtained by clustering, where the number of VMD decompositions is obtained according to the calculated center frequency; Step 5: Train the load trend informer model: divide the data set, construct the input of the informer model according to the sliding window method, and predict the charging trend curve; Step 6: Load curve shape correction: Replace the load peak point in the charging load curve of the target residential area predicted by the Informer model with the peak value predicted by the LSTM model in step 3; Step 7: Model evaluation: By calculating the values ​​of different error evaluation indicators under different prediction scenarios, the prediction accuracy of the model is comprehensively evaluated.

2. According to claim 1, a method for predicting charging pile load in residential areas based on deep learning and load peak correction is characterized in that: The specific method of data preprocessing is: 1) Identification and correction of outliers: Combine the historical charging behavior of charging pile users and the capacity of the charging pile, set the maximum load and the longest and lowest charging time thresholds, and mark the abnormal data for charging behavior data that exceeds or is significantly lower than this range for correction. The outliers are corrected by filling the mean. The specific formula is as follows: Where: y(k,t) represents the charging load value of the target residential area at time t on the kth day; 2) Missing value filling: In order to maintain the continuity of the time series, the missing values ​​in the corrected data are filled according to the charging load of the adjacent charging date or similar day. The specific formula is as follows: Y(d,t)=ω1Y(d,t-1)+ω2Y(d,t+1)+ω3Y(d-1,t)+ω4Y(d+1,t) Where: Y(d,t) represents the charging load value of the target residential area at time t on the dth day; ω is the weight coefficient of different similar days; 3) Standardization: In the process of charging load data standardization, the maximum and minimum normalization method is selected. The data of different dimensions are scaled according to a certain ratio and mapped to [0,1]. The advantage of maximum and minimum normalization is that the distribution form of the original data is retained and it is suitable for situations where data needs to be mapped to a specific range. The calculation formula of maximum and minimum normalization is: Where: X' is the normalized charging load; X represents the original load data; X max , X min Respectively represent the maximum and minimum values ​​in the load sequence; When the prediction model is trained, the predicted value it outputs is still in the interval [0,1]. At this time, the predicted value needs to be denormalized to restore it to its initial order of magnitude. The denormalization calculation formula is as follows: X=X′*(X max -X min )+X min Where: X represents the load data after denormalization; X′ represents the predicted load value initially obtained by LSTM; X max ,X min Represent the maximum and minimum values ​​in the data sequence respectively.

3. The method for predicting charging pile load in residential areas based on deep learning and load peak correction according to claim 1, characterized in that: The specific method of feature selection is: For the daily charging load peak of the target residential area, MIC is used to calculate the correlation between the number of charging piles installed in the residential area on the corresponding date, the average temperature, air pressure, precipitation, electricity price, season, whether it is a holiday, and other influencing features and the daily load peak. For the charging load trend curve with 15 minutes as the sampling point, MIC is used to calculate the correlation between the number of charging piles that are charging and starting to charge and the load at each sampling point, and select the features with the strongest correlation as the input variables of the prediction model to improve the prediction accuracy and reduce the model calculation complexity. The detailed calculation formula of MIC is as follows: Where: a and b are the interval numbers in the x and y directions respectively; s is a variable whose size is equal to 0.6 times the amount of data; p(x,y) is the joint probability distribution of x and y; p(x) and p(y) are the marginal probability distributions of x and y respectively; the MIC value ranges from 0 to 1, and the larger the value, the stronger the correlation between the two variables.

4. According to claim 1, a method for predicting charging pile load in residential areas based on deep learning and load peak correction is characterized in that: The specific steps of using LSTM to predict daily load peak are as follows: To predict multivariate loads using the LSTM model, we first confirm the input time series data length of the load peak prediction model, including the historical load peak and the time series length of the influencing factors, and then set the structure of the LSTM prediction model, including the optimal number of layers, number of neurons, optimizer, and activation function type of the model.

5. The method for predicting charging pile load in residential areas based on deep learning and load peak correction according to claim 1 is characterized in that: The specific steps of decomposing and aggregating the load trend curve are as follows: The charging load sequence of the target residential area is adaptively decomposed using fully integrated empirical mode decomposition with adaptive noise (CEEMDAN) to obtain multiple subsequences, and the sample entropy of each subsequence is calculated respectively. The complexity of each subsequence is evaluated by the size of the sample entropy value. Then, the K-Means clustering algorithm is used to cluster the sample entropy values ​​of each subsequence decomposed by CEEMDAN. By setting the clustering parameter of K-Means to 3, the corresponding subsequences are aggregated into low, medium, and high frequency components respectively. Distinguishing components of different complexity helps to improve the accuracy of prediction.

6. The method for predicting charging pile load in residential areas based on deep learning and load peak correction according to claim 1, characterized in that: The specific steps of the secondary decomposition of the load trend curve are as follows: Since the high frequency after aggregation is still complex and difficult to predict, the variational mode decomposition (VMD) decomposition algorithm is used to perform secondary decomposition of the high frequency components, and the corresponding central frequencies for different decomposition numbers K are calculated respectively. When the central frequency remains stable as the K value increases, the K value is determined to be the optimal decomposition number of the VMD algorithm; the K subsequences, a total of K+2 subsequences, obtained by secondary decomposition of the low-frequency components, medium-frequency components and high-frequency components aggregated in claim 4 are combined and reconstructed as input for load trend prediction.

7. A method for predicting charging pile load in residential areas based on deep learning and load peak correction according to claim 1, characterized in that The Informer model is used to predict load trends. The specific steps are as follows: Compared with other deep learning and machine learning models, Informer is better at obtaining long-term time dependencies in load data. It uses a sliding window method to sequentially take data sets with time series characteristics as the input of the model, and performs weighted superposition of the prediction results of each model through a fully connected layer to obtain the final charging load trend prediction result.

8. The method for predicting charging pile load in residential areas based on deep learning and load peak correction according to claim 1 is characterized in that: The Adam algorithm is used to optimize the training model. The specific steps are as follows: The Adam algorithm is selected to optimize the Informer model. During the model training process, the learning rate of the Adam algorithm will be adaptively adjusted according to the model training situation. It is a further improvement on the gradient descent method, making the model more accurate during the training process.

9. The method for predicting charging pile load in residential areas based on deep learning and load peak correction according to claim 1, characterized in that: Load peak curve correction, the specific steps are as follows: The charging load trend curve can reflect the total load demand of charging piles in the target residential area in the future period, but the peak prediction is not accurate enough. In order to improve the accuracy of the daily load peak prediction and ensure the safety of transformers in residential areas, the load peak in the predicted load curve is replaced with the target residential area charging daily load peak predicted by LSTM to obtain the final charging load prediction value of the target residential area.

10. The method for predicting charging pile load in residential areas based on deep learning and load peak correction according to claim 1, characterized in that: Different error indicators are used to evaluate the prediction performance of the model and the accuracy of the model's charging load prediction results is evaluated.

Images