Electric vehicle charging station load prediction method based on ANVD decomposition and typical user characteristic DTW matching

Abstract

The invention discloses an electric vehicle charging station load prediction method based on ANVD decomposition and typical user feature DTW matching, and the method comprises the steps: carrying out the decomposition of an electric vehicle charging station load time sequence through ANVD self-adaptive non-negative variation decomposition, and obtaining a charging load IMF; according to the electric vehicle charging station load time sequence, typical charging characteristics of electric vehicle users are constructed; carrying out classification matching based on DTW dynamic time planning on each charging load IMF and the constructed typical charging characteristics to obtain different types of characteristic time sequences; constructing a charging load prediction model based on a TCN time domain convolutional network, and training and predicting the charging load prediction model for different types of feature time sequences; and superposing prediction results corresponding to all category feature time sequences to obtain an electric vehicle charging station load prediction result. According to the method, ANVD decomposition and classification matching based on DTW dynamic time planning are applied to electric vehicle charging station load prediction at the same time, and the accuracy of the electric vehicle charging station load prediction model is effectively improved.

Description

Technical Field

[0001] The present invention belongs to the field of new energy, relates to a load prediction technology for a vehicle charging station, and specifically relates to a load prediction method for a vehicle charging station based on ANVD decomposition and DTW matching of typical user characteristics. Background Art

[0002] Against the backdrop of global warming, energy shortages, environmental protection needs, and rapid technological development, electric vehicles have been actively promoted by governments and enterprises around the world. my country's electric vehicle charging infrastructure has achieved leapfrog development, charging technology has been rapidly improved, the standard system has been gradually improved, and the industrial ecology has been steadily formed. The world's largest charging infrastructure system with the largest radiation area and the most comprehensive service vehicles has been built. However, the rapid development has led to an increasing charging load, the complexity of different types of electric vehicle charging loads, and the fluctuations in electric vehicle charging loads due to different weather and charging habits of different users will have an impact on the power grid that cannot be ignored. Therefore, accurate prediction of electric vehicle charging load is an important part of building a sustainable electric vehicle industry and intelligent energy management.

[0003] In the past few years, many researchers have actively explored various methods to predict the charging load of electric vehicles. At present, the mainstream methods for electric vehicle load forecasting mainly include statistical models, machine learning, optimization and hybrid methods. Statistical models use historical data and statistical analysis to predict loads. Time series analysis, regression models, etc. can be used to establish prediction models. Machine learning methods use historical data to train models and make predictions by learning the patterns and relationships of data. Common algorithms include support vector machines, random forests and neural networks. Optimization methods transform load forecasting problems into optimization problems, consider factors such as charging demand and energy costs, and solve the optimal charging strategy through optimization algorithms. Hybrid methods combine the advantages of multiple methods, use statistical models to deal with long-term trends and seasonal effects, and use machine learning methods to deal with complex nonlinear relationships.

[0004] Existing machine learning electric vehicle charging load forecasting methods have the following problems: First, they rely on historical data and ignore the complex nonlinear characteristics and randomness in electric vehicle charging behavior, especially the behavioral differences between different users. Second, these methods are usually difficult to deal with problems such as strong load volatility and low temporal resolution, resulting in low prediction accuracy. In addition, some traditional machine learning methods such as support vector machines and random forests are susceptible to overfitting problems when processing high-dimensional time series data, and the model has poor interpretability, making it difficult to effectively capture the variable charging load characteristics. Summary of the invention

[0005] Purpose of the invention: In order to overcome the deficiencies in the prior art, a load forecasting method for a vehicle charging station based on ANVD decomposition and DTW matching of typical user characteristics is provided to improve the accuracy of the load forecasting model for a vehicle charging station.

[0006] Technical solution: To achieve the above purpose, the present invention provides a method for predicting the load of a vehicle charging station based on ANVD decomposition and DTW matching of typical user characteristics, comprising the following steps:

[0007] S1: ANVD adaptive non-negative variational decomposition is used to decompose the load time series of the electric vehicle charging station to obtain the charging load IMF;

[0008] S2: Construct typical charging characteristics of electric vehicle users based on the load sequence of electric vehicle charging stations;

[0009] S3: Classify and match each charging load IMF decomposed in step S1 with the typical charging features constructed in step S2 based on DTW dynamic time planning to obtain feature time series of different categories;

[0010] S4: A charging load prediction model is constructed based on the TCN time domain convolutional network, and the charging load prediction model is trained and predicted for different categories of feature time series;

[0011] S5: The prediction results corresponding to all category feature time series are superimposed to obtain the load prediction result of the electric vehicle charging station.

[0012] Furthermore, the step S1 is specifically as follows:

[0013] ANVD (adaptive non-negative variational decomposition, ANVD) uses the alternating direction multiplier method ADMM for optimization based on VMD variational mode decomposition, introduces penalty terms and Lagrangian multipliers λ(t), and forms the following Lagrangian function based on the VMD objective function:

[0014]

[0015] In the formula, u k are the decomposed modes; ω k is the frequency set corresponding to the modal function; δ(t) is the Dirac function; α is the regularization parameter, which controls the weight of the regularization term; λ is the Lagrange multiplier, which is used for the dual update in the ADMM algorithm; f(t) is the original load time series.

[0016] Furthermore, in step S1, to ensure that each charging load component u k are all non-negative, and non-negativity constraints need to be added to the optimization problem:

[0017]

[0018] Therefore, the optimization objective becomes to minimize the Lagrangian function under the condition of non-negativity:

[0019]

[0020] In the optimization process of ADMM, when updating each modal function u k When , it is necessary to ensure its non-negativity, and the formula is:

[0021]

[0022] In the formula, represents the real part; F and F -1 denote Fourier transform and inverse Fourier transform respectively;

[0023] The updates of the center frequency and Lagrange multipliers are not affected by the non-negativity constraint and are performed according to the standard VMD update formula.

[0024] Furthermore, the step S2 is specifically as follows:

[0025] Collect charging information of private trams, online ride-hailing vehicles, public buses, etc. at charging stations, and extract characteristic indicators of charging information, including: charging frequency, charging time, charging power, daily charging volume, and charging time period distribution;

[0026] According to the extracted characteristic indicators, the Monte Carlo sampling method is used to simulate the charging behavior of different types of tram users. A large number of possible charging behavior samples are generated from historical data through random sampling, and based on the results of Monte Carlo simulation, a typical charging characteristic curve is generated for each type of tram user.

[0027] Furthermore, in step S3, the decomposed IMFs are matched with the constructed typical charging characteristics by classification based on DTW dynamic time planning. Considering the problem that the constructed typical characteristic curve does not match the time resolution of IMF, DTW can align the two time series in nonlinear time by stretching or compressing the time axis, so as to find the best matching path between them;

[0028] For each decomposed charging load IMF, it is compared with all typical user charging features. The DTW (dynamic timewarping, DTW) algorithm is used to calculate the DTW distance between each IMF and the typical feature to find the best matching path. Then, using the DTW distance as a similarity metric, each IMF is classified into the most similar typical user charging feature category for subsequent training and prediction.

[0029] Furthermore, the step S3 is specifically as follows:

[0030] For the IMF sequence X=(x1,x2,…,x n ) and the typical charging characteristic sequence Y=(y1,y2,…,y m ), calculate the Euclidean distance matrix D between two sequences, where each element d i,j The calculation formula is:

[0031] d i,j =(x i -y j ) 2

[0032] Define the cumulative distance matrix C to store the minimum cumulative distance from the starting point to any point; the initial conditions are:

[0033] C(1,1)=d 1,1

[0034] The recursive formula is:

[0035] C(i,j)=d i,j +min{C(i-1,j),C(i,j-1),C(i-1,j-1)}

[0036] Starting from C(n,m), trace the minimum path in reverse until reaching C(1,1), and obtain the best matching path W = (w1,w2,...,w k ), where each w k =(i k ,j k ) represents the matching points in sequences X and Y;

[0037] The DTW distance is defined as the value of the final position in the cumulative distance matrix, which represents the minimum matching cost of two sequences:

[0038] DTW(X,Y)=C(n,m)

[0039] By calculating the DTW distance between each IMF and each typical user charging feature, the IMF is classified into the most similar feature category and the feature time series of different categories are obtained.

[0040] Furthermore, the TCN time-domain convolutional network in step S4 includes causal convolution, dilated convolution, residual connection and loss function, which are as follows:

[0041] Causal convolution: ensures that the current output depends only on the current and previous inputs to avoid information leakage; if the convolution kernel is k, then the output at time step t depends only on the input moments t, t-1, ..., t-k+1, as shown in the following formula:

[0042]

[0043] Among them, w i is the weight of the convolution kernel, x t-i is the corresponding input;

[0044] Atrous convolution: Add a "atrous" structure to causal convolution, that is, skip a certain input interval during the convolution process to expand the receptive field; the formula for atrous convolution is:

[0045]

[0046] Among them, r is the void ratio;

[0047] Residual connection: To solve the gradient vanishing problem in deep network training, TCN introduces residual connection to add the input directly to the output; the formula of the residual module is:

[0048]

[0049] Wherein, σ is an activation function, and the present invention adopts the ReLU activation function;

[0050] TCN uses mean square error MSE as the loss function; for time series prediction problems, let is the predicted value of the model at time t, and the true value is y t , then the loss function is defined as:

[0051]

[0052] Where N is the number of samples.

[0053] Furthermore, the training of the charging load prediction model in step S4 uses a back propagation algorithm to perform gradient descent to update parameters, and the specific steps of the training process are as follows:

[0054] 1) Initialize model parameters: Initialize the convolution kernel weight w and bias term b using random initialization method;

[0055] 2) Forward propagation: The input data passes through the causal convolution and the dilated convolution layer, and the output at each time t is calculated. For each time, the loss function MSE is calculated, and the difference between the output result and the true value is calculated;

[0056] 3) Back propagation: Calculate the gradient according to the loss function and Where L is the total loss; the gradient with respect to weight w is:

[0057]

[0058] 4) Parameter update: Use the gradient descent method to update the weights and bias terms. The formula is:

[0059]

[0060] Update the bias term:

[0061]

[0062] Where η is the learning rate;

[0063] 5) Through multiple iterative training, the TCN network gradually learns the characteristics of time series data and reduces the prediction error.

[0064] Furthermore, the method for predicting the charging load prediction model in step S4 is:

[0065] The newly input time series data, that is, the IMF feature data of different time periods, is input into the trained TCN model. The convolution layer of the TCN model processes the input data in sequence, obtains the time series characteristics of the charging load through causal convolution and void convolution, and outputs the predicted charging load value at each moment. For each time series input data, the model will give the predicted load value of the corresponding time period.

[0066] The TCN model combines the advantages of one-dimensional convolutional neural network (1D-CNN) and recurrent neural network (RNN) and is specifically designed for time series data prediction tasks.

[0067] Through DTW classification matching, different frequency components (IMF) can be effectively matched and classified with typical charging characteristics, so that when building a load forecasting model in the time domain convolutional network (TCN), targeted training and prediction can be performed according to the feature category.

[0068] Beneficial effects: Compared with the prior art, the present invention, by introducing the ANVD decomposition technology, can decompose the load time series data into multiple modal components of different frequencies, ensuring that the decomposition results are non-negative, thereby more accurately reflecting the load change law of the electric vehicle charging station. By constructing the charging characteristics of typical users, it can better adapt to the differences in charging behaviors of different types of users in different time periods. Combined with the DTW dynamic time planning algorithm, the decomposed mode is matched with the user's charging characteristics to effectively capture the load fluctuation law and adapt to the charging behavior of different types of users. The load forecasting model of the TCN time domain convolutional network combines the advantages of one-dimensional convolutional neural network (1D-CNN) and recurrent neural network (RNN), and can more effectively process complex time series data, further improving the prediction effect.

[0069] In addition, the present invention also has the advantages of strong generalization ability and high feature extraction efficiency. By generating typical charging characteristics of different types of tram users through Monte Carlo sampling, the model can better adapt to the diverse user behaviors, and combined with the TCN time domain convolutional network to effectively process complex nonlinear time series data, further improving the stability and reliability of load forecasting. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] Figure 1 is a flow chart of the prediction method of the present invention;

[0071] Figure 2 It is the ANVD adaptive non-negative variational decomposition graph;

[0072] Figure 3 This is a typical charging characteristic curve for electric vehicle users;

[0073] Figure 4 It is the charging load prediction curve of the electric vehicle charging station. DETAILED DESCRIPTION

[0074] The present invention is further explained below in conjunction with the accompanying drawings and specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention. After reading the present invention, various equivalent forms of modifications to the present invention by those skilled in the art all fall within the scope defined by the claims attached to this application.

[0075] like Figure 1 As shown, the present invention provides a method for predicting the load of a vehicle charging station based on ANVD decomposition and DTW matching of typical user characteristics, comprising the following steps:

[0076] S1: ANVD adaptive non-negative variational decomposition is used to decompose the load time series of the electric vehicle charging station to obtain the charging load IMF:

[0077] Based on the variational mode decomposition of VMD, ANVD uses the alternating direction multiplier method ADMM for optimization, introduces penalty terms and Lagrangian multipliers λ(t), and forms the following Lagrangian function based on the VMD objective function:

[0078]

[0079] In the formula, u k are the decomposed modes; ω k is the frequency set corresponding to the modal function; δ(t) is the Dirac function; α is the regularization parameter, which controls the weight of the regularization term; λ is the Lagrange multiplier, which is used for the dual update in the ADMM algorithm; f(t) is the original load time series.

[0080] To ensure that each charging load component u kare all non-negative, and non-negativity constraints need to be added to the optimization problem:

[0081]

[0082] Therefore, the optimization objective becomes to minimize the Lagrangian function under the condition of non-negativity:

[0083]

[0084] In the optimization process of ADMM, when updating each modal function u k When , it is necessary to ensure its non-negativity, and the formula is:

[0085]

[0086] In the formula, represents the real part; F and F -1 denote Fourier transform and inverse Fourier transform respectively;

[0087] The updates of the center frequency and Lagrange multipliers are not affected by the non-negativity constraint and are performed according to the standard VMD update formula.

[0088] S2: Based on the load sequence of the electric vehicle charging station, construct the typical charging characteristics of electric vehicle users:

[0089] Collect charging information of private trams, online ride-hailing vehicles, public buses, etc. at charging stations, and extract characteristic indicators of charging information, including: charging frequency, charging time, charging power, daily charging volume, and charging time period distribution;

[0090] According to the extracted characteristic indicators, the Monte Carlo sampling method is used to simulate the charging behavior of different types of tram users. A large number of possible charging behavior samples are generated from historical data through random sampling, and based on the results of Monte Carlo simulation, a typical charging characteristic curve is generated for each type of tram user.

[0091] S3: Classify and match each charging load IMF decomposed in step S1 with the typical charging features constructed in step S2 based on DTW dynamic time planning to obtain feature time series of different categories:

[0092] For the IMF sequence X=(x1,x2,...,x n ) and the typical charging characteristic sequence Y=(y1,y2,...,y m ), calculate the Euclidean distance matrix D between two sequences, where each element d i,j The calculation formula is:

[0093] d i,j =(x i -y j )2

[0094] Define the cumulative distance matrix C to store the minimum cumulative distance from the starting point to any point; the initial conditions are:

[0095] C(1,1)=d 1,1

[0096] The recursive formula is:

[0097] C(i,j)=d i,j +min{C(i-1,j),C(i,j-1),C(i-1,j-1)}

[0098] Starting from C(n,m), trace the minimum path in reverse until reaching C(1,1), and obtain the best matching path W = (w1,w2,…,w k ), where each w k =(i k ,j k ) represents the matching points in sequences X and Y;

[0099] The DTW distance is defined as the value of the final position in the cumulative distance matrix, which represents the minimum matching cost of two sequences:

[0100] DTW(X,Y)=C(n,m)

[0101] By calculating the DTW distance between each IMF and each typical user charging feature, the IMF is classified into the most similar feature category and the feature time series of different categories are obtained.

[0102] S4: A charging load prediction model is constructed based on the TCN time domain convolutional network, and the charging load prediction model is trained and predicted for different categories of feature time series;

[0103] The TCN time-domain convolutional network combines the advantages of one-dimensional convolution and causal convolution, and can effectively handle long temporal dependency problems.

[0104] Causal convolution: ensures that the output at the current moment depends only on the current and previous inputs to avoid information leakage. If the convolution kernel is k, then the output at time step t depends only on the input moments t, t-1, ..., t-k+1, and the formula is as follows:

[0105]

[0106] Among them, w i is the weight of the convolution kernel, x t-i is the corresponding input.

[0107] Atrous convolution: Adding a "atrous" structure to causal convolution means skipping a certain input interval during the convolution process, thereby expanding the receptive field. The formula for atrous convolution is:

[0108]

[0109] Where r is the void ratio.

[0110] Residual Connections: To solve the gradient vanishing problem in deep network training, TCN introduces residual connections to add the input directly to the output. The formula of the residual module is:

[0111]

[0112] Wherein, σ is an activation function, and the present invention adopts the ReLU activation function.

[0113] TCN uses mean square error (MSE) as the loss function. For the time series prediction problem, let is the predicted value of the model at time t, and the true value is y t , then the loss function is defined as:

[0114]

[0115] Where N is the number of samples.

[0116] The construction of the charging load prediction model is as follows:

[0117] ANVD decomposition: First, adaptive non-negative variational decomposition (ANVD) is performed on the load time series data of the electric vehicle charging station to decompose it into multiple intrinsic mode functions (IMFs), each of which represents the load characteristics at different frequencies.

[0118] Construction of typical charging characteristics of users: Collect charging data of users at charging stations and extract characteristics such as charging frequency, time, and power. Then simulate the charging behavior of different types of tram users (such as private trams, online ride-hailing cars, and public buses) through the Monte Carlo sampling method to generate a typical charging characteristic curve.

[0119] DTW classification matching: Each IMF is matched with the typical charging features based on dynamic time planning (DTW), the DTW distance between the two is calculated, and the IMF is classified into the corresponding typical user category according to the similarity.

[0120] Build a prediction model: Use the time-domain convolutional network (TCN) to train and predict the model for different categories of feature time series.

[0121] Superimpose prediction results: Finally, superimpose the prediction results of all category feature time series to obtain the final charging load prediction value.

[0122] The charging load prediction model is trained and predicted according to different categories of feature time series. The training method is:

[0123] TCN training uses the backpropagation algorithm to perform gradient descent to update parameters. The specific steps of the training process are as follows:

[0124] Step 1: Initialize model parameters: Initialize the convolution kernel weight w and bias term b using random initialization method.

[0125] Step 2, forward propagation: The input data passes through the causal convolution and the dilated convolution layer, and the output at each time t is calculated. For each time, the loss function MSE is calculated, and the difference between the output result and the true value is calculated.

[0126] Step 3, back propagation: calculate the gradient according to the loss function and Where L is the total loss. The gradient for weight w is:

[0127]

[0128] Step 4, parameter update: Use the gradient descent method to update the weights and bias terms. The formula is:

[0129]

[0130] Update the bias term:

[0131]

[0132] Where η is the learning rate.

[0133] Step 5: Through multiple iterative training, the TCN network gradually learns the characteristics of time series data and reduces the prediction error.

[0134] The prediction method is:

[0135] After training is completed, the TCN model will perform load forecasting based on new input data. Although TCN does not have an explicit mathematical expression like a traditional regression model, its core operation can be expressed as: input the newly input time series data (i.e., IMF feature data in different time periods) into the trained TCN model. The convolutional layer of the TCN model processes the input data in sequence, and obtains the time series characteristics of the charging load through causal convolution and void convolution. The predicted charging load value is output at each moment. For each time series input data, the model will give the predicted load value for the corresponding time period.

[0136] S5: The prediction results corresponding to all category feature time series are superimposed to obtain the load prediction result of the electric vehicle charging station.

[0137] Based on the above scheme, in order to verify the effectiveness of the method of the present invention, this embodiment is verified by experiments, as follows:

[0138] The experimental part uses the historical charging load data of an electric vehicle charging station in Jiangbei, Nanjing, Jiangsu Province. The historical load data is from 0:00 on July 14, 2022 to 0:00 on December 14, 2022, with a sampling frequency of 15min / point and the data unit is kW.

[0139] Firstly, the ANVD adaptive non-negative variational decomposition is used to decompose the load time series of the electric vehicle charging station. The decomposition diagram is shown in Figure 2 As shown. Figure 2 It can be seen that the ANVD adaptive non-negative variational decomposition decomposes the original load curve into 9 non-negative load IMFs, which represent the load characteristics under different time and frequency.

[0140] Next, we collect charging information of private trams, online ride-hailing cars, public buses, etc. at charging stations, and extract characteristic indicators including charging frequency, charging time, charging power, daily charging amount, and charging time distribution. We then use the Monte Carlo sampling method to simulate the charging behavior of different types of tram users. Through the random sampling method, we generate a large number of possible charging behavior samples from historical data, and based on the results of the Monte Carlo simulation, we generate a typical charging characteristic curve for each type of tram user, such as Figure 3 shown.

[0141] according to Figure 3 , the charging behavior of private tram charging characteristic curve is concentrated between 20:00 and 23:00 in the evening, and the peak occurs at around 21:00, which means that private trams are usually charged after get off work or at night. The charging behavior of the online car-hailing charging characteristic curve is more scattered, with two obvious peaks, which appear at lunch time (13:00) and evening break time (22:00), which reflects the characteristics of online car-hailing charging during work breaks. Public bus charging characteristic curve: The charging behavior is mainly concentrated between 1:00 and 4:00 at night, and the peak occurs at around 2:00, indicating that public buses are usually charged in batches after the night operation ends.

[0142] In the prediction stage, a charging load prediction model based on the TCN time domain convolutional network is built for various feature time series for training and prediction.

[0143] In the comparative experiment, the LSTM prediction model, the BiLSTM prediction model and the TCN prediction model provided by the present invention are introduced for comparative experiments, and then the difference between the proposed method and other methods is studied. Figure 4 As shown. Figure 4 It can be seen that the prediction curve of the TCN prediction model of the present invention is closest to the true value and has the best prediction effect. It can be seen that ANVD and feature matching play an important role in improving the prediction effect. It can identify typical user characteristics and improve the learning ability of the neural network for different charging characteristics, thereby improving the accuracy of the prediction.

[0144] Table 1 Evaluation indicators of prediction results

[0145]

[0146] According to the evaluation index data table 1, the ANVD-FCM model proposed in this invention has the lowest RMSE and MAE, which are 0.4534% and 0.3792% respectively, and its prediction effect is the best. The second is the BiLSTM model, with an RMSE of 1.0034% and a MAE of 0.8391%. The single LSTM model has the highest RMSE and MAE, which are 3.2456% and 2.7142% respectively, and its prediction effect is not good.

[0147] In summary, based on the given evaluation index data, it can be concluded that the ANVD-FCM model proposed in the present invention performs best and has the best prediction effect, which verifies the effectiveness and practical effect of the method of the present invention.

Claims

1. A method for predicting the load of electric vehicle charging stations based on ANVD decomposition and DTW matching of typical user characteristics, characterized in that: The steps include: S1: ANVD adaptive non-negative variational decomposition is used to decompose the load time series of the electric vehicle charging station to obtain the charging load IMF; S2: Construct typical charging characteristics of electric vehicle users based on the load sequence of electric vehicle charging stations; S3: Classify and match each charging load IMF decomposed in step S1 with the typical charging features constructed in step S2 based on DTW dynamic time planning to obtain feature time series of different categories; S4: A charging load prediction model is constructed based on the TCN time domain convolutional network, and the charging load prediction model is trained and predicted for different categories of feature time series; S5: The prediction results corresponding to all category feature time series are superimposed to obtain the load prediction result of the electric vehicle charging station.

2. According to claim 1, a method for predicting the load of a vehicle charging station based on ANVD decomposition and DTW matching of typical user characteristics is characterized in that: The step S1 is specifically as follows: Based on the variational mode decomposition of VMD, ANVD uses the alternating direction multiplier method ADMM for optimization, introduces penalty terms and Lagrangian multipliers λ(t), and forms the following Lagrangian function based on the VMD objective function: In the formula, u k are the decomposed modes; ω k is the frequency set corresponding to the modal function; δ(t) is the Dirac function; α is the regularization parameter, which controls the weight of the regularization term; λ is the Lagrange multiplier used for dual update in the ADMM algorithm; f(t) is the original load time series.

3. According to claim 2, a method for predicting the load of a vehicle charging station based on ANVD decomposition and DTW matching of typical user characteristics is characterized in that: In step S1, each charging load component u k are all non-negative, and non-negativity constraints need to be added to the optimization problem: Therefore, the optimization goal becomes to minimize the Lagrangian function under the condition of non-negativity: In the optimization process of ADMM, when updating each modal function u k When , it is necessary to ensure its non-negativity, and the formula is: In the formula, R represents the real part; F and F -1 denote Fourier transform and inverse Fourier transform respectively; The updates of the center frequency and Lagrange multipliers are not affected by the non-negativity constraint and are performed according to the standard VMD update formula.

4. According to claim 1, a method for predicting the load of a vehicle charging station based on ANVD decomposition and DTW matching of typical user characteristics is characterized in that: The step S2 is specifically as follows: Collect charging information and extract characteristic indicators of charging information, including: charging frequency, charging time, charging power, daily charging amount and charging time period distribution; According to the extracted characteristic indicators, the Monte Carlo sampling method is used to simulate the charging behavior of different types of tram users. A large number of possible charging behavior samples are generated from historical data through random sampling, and based on the results of Monte Carlo simulation, a typical charging characteristic curve is generated for each type of tram user.

5. According to claim 1, a method for predicting the load of a vehicle charging station based on ANVD decomposition and DTW matching of typical user characteristics is characterized in that: In the step S3, for each decomposed charging load IMF, it is compared with all typical user charging features; the DTW distance between each IMF and the typical feature is calculated by the DTW algorithm to find the best matching path; then, using the DTW distance as a similarity measurement indicator, each IMF is classified into the most similar typical user charging feature category for subsequent training and prediction.

6. The method for predicting the load of a vehicle charging station based on ANVD decomposition and DTW matching of typical user characteristics according to claim 5 is characterized in that: The step S3 is specifically as follows: For the IMF sequence X=(x1,x2,...,x n ) and the typical charging characteristic sequence Y=(y1,y2,...,y m ), calculate the Euclidean distance matrix D between two sequences, where each element d i,j The calculation formula is: d i,j =(x i -y j ) 2 Define the cumulative distance matrix C to store the minimum cumulative distance from the starting point to any point; the initial conditions are: C(1,1)=d 1,1 The recursive formula is: C(i,j)=d i,j +min{C(i-1,j),C(i,j-1),C(i-1,j-1)} Starting from C(n,m), trace the minimum path in reverse until reaching C(1,1), and obtain the best matching path W = (w1,w2,...,w k ), where each w k =(i k ,j k ) represents the matching points in sequences X and Y; The DTW distance is defined as the value of the final position in the cumulative distance matrix, which represents the minimum matching cost of two sequences: DTW(X,Y)=C(n,m) By calculating the DTW distance between each IMF and each typical user charging feature, the IMF is classified into the most similar feature category and the feature time series of different categories are obtained.

7. The method for predicting the load of a vehicle charging station based on ANVD decomposition and DTW matching of typical user characteristics according to claim 1 is characterized in that: The TCN time-domain convolutional network in step S4 includes causal convolution, dilated convolution, residual connection and loss function, which are as follows: Causal convolution: ensures that the current output depends only on the current and previous inputs to avoid information leakage; if the convolution kernel is k, then the output at time step t depends only on the input moments t, t-1, ..., t-k+1, as shown in the following formula: Among them, w i is the weight of the convolution kernel, x t-i is the corresponding input; Atrous convolution: Add a "atrous" structure to causal convolution, that is, skip a certain input interval during the convolution process to expand the receptive field; the formula for atrous convolution is: Among them, r is the void ratio; Residual connection: To solve the gradient vanishing problem in deep network training, TCN introduces residual connection to add the input directly to the output; the formula of the residual module is: Wherein, σ is an activation function, and the present invention adopts the ReLU activation function; TCN uses mean square error MSE as the loss function; for time series prediction problems, let is the predicted value of the model at time t, and the true value is y t , then the loss function is defined as: Where N is the number of samples.

8. The method for predicting the load of a vehicle charging station based on ANVD decomposition and DTW matching of typical user characteristics according to claim 7 is characterized in that: The training of the charging load prediction model in step S4 adopts the back propagation algorithm to perform gradient descent to update parameters. The specific steps of the training process are as follows: 1) Initialize model parameters: Initialize the convolution kernel weight w and bias term b using random initialization method; 2) Forward propagation: The input data passes through the causal convolution and the dilated convolution layer, and the output at each time t is calculated. For each time, the loss function MSE is calculated, and the difference between the output result and the true value is calculated; 3) Back propagation: Calculate the gradient according to the loss function and Where L is the total loss; the gradient with respect to weight w is: 4) Parameter update: Use the gradient descent method to update the weights and bias terms. The formula is: Update the bias term: Where η is the learning rate; 5) Through multiple iterative training, the TCN network gradually learns the characteristics of time series data and reduces the prediction error.

9. The method for predicting the load of a vehicle charging station based on ANVD decomposition and DTW matching of typical user characteristics according to claim 1 is characterized in that: The method for predicting the charging load prediction model in step S4 is: The newly input time series data, that is, the IMF feature data of different time periods, is input into the trained TCN model. The convolution layer of the TCN model processes the input data in sequence, obtains the time series characteristics of the charging load through causal convolution and void convolution, and outputs the predicted charging load value at each moment. For each time series input data, the model will give the predicted load value of the corresponding time period.

Images