An electric vehicle charging load prediction method based on an improved time convolution network

Abstract

The application provides an electric vehicle charging load prediction method based on an improved time convolution network, and relates to the technical field of electric vehicle charging load prediction.The method decomposes the electric vehicle charging load time series into relatively simple components through variational mode decomposition (VMD).In order to avoid the calculation being complicated and the error being accumulated due to too many components, firstly, the complexity of each component is calculated by using sample entropy, and the components are superimposed and combined to obtain the trend component, the long-time scale detail component, the short-time scale detail component and the fluctuation component of the electric vehicle charging load.Then, a double-end attention mechanism is added on the basis of the time convolution network (TCN).Finally, an Attention-TCN prediction model is established for each component, the prediction results of the components are reconstructed, and then the multi-step prediction of the electric vehicle charging load is realized.The accuracy of the electric vehicle charging load under the multi-step prediction condition is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of electric vehicle charging load prediction, and particularly relates to an electric vehicle charging load prediction method based on an improved time convolution network. BACKGROUND

[0002] With the wide popularity of electric vehicles (EVs), the electrical energy stored in the batteries provides the possibility for demand response and grid support in extreme situations. However, the high popularity of EVs also brings many adverse effects to the power grid. The load prediction accuracy of the charging station, as the connection point of the EV and the power grid, can not only ensure the economic and stable operation of the power system, but also has an important role in the peak shaving and frequency modulation functions of the power system. At the same time, as a power grid resource, the use of intelligent charging to improve the utilization efficiency of EVs also cannot do without accurate load prediction. Therefore, more accurate EV charging load prediction results are crucial for supporting the development of EV charging, promoting the efficient use of EVs, and ensuring the safety of the power grid.

[0003] As a new type of load, the load characteristics of electric vehicle charging are influenced by meteorological factors, date types, traffic conditions, and user travel behavior rules, and therefore it is difficult to predict the electric vehicle charging load. Based on this, in the document “Wang Yi, Gu Yi, Ding Zhuang, etc. Electric vehicle charging demand prediction based on EMD-fuzzy entropy and ensemble learning. Power System Automation, 2020, 44(3): 114-121”, the EMD decomposition algorithm is applied to decompose the electric vehicle charging load demand into a series of modal functions, then the fuzzy entropy of the components is calculated to reduce the number of modal components, and then the SVR and LSTM models are applied for prediction, and the prediction results of each component are combined to obtain the electric vehicle charging load demand. However, the EMD decomposition method has a significant modal aliasing phenomenon, and therefore cannot fully reflect the change rule of the electric vehicle charging load at different time scales, thereby affecting the accuracy of the prediction model of each component. In addition, recursive models such as LSTM and GRU, although they contain time memory units, still have the problem of forgetting historical information, and the ability to mine potential relationships between non-continuous data is weak.

[0004] Therefore, it is a technical problem to be solved by those skilled in the art to design an electric vehicle charging load prediction method based on an improved time convolution network to solve the above problems. SUMMARY

[0005] The technical problem to be solved by the present application is to overcome the deficiencies in the prior art and provide an electric vehicle charging load prediction method based on an improved time convolution network.

[0006] The application is implemented by the following technical scheme: a charging load prediction method for electric vehicles based on an improved time convolution network, comprising the following steps:

[0007] S1: Obtain electric vehicle charging load data, and fill in missing values, that is, fill in missing values by using the average of data of adjacent two time points;

[0008] S2: Take the time series of the electric vehicle charging load data obtained in S1 as the original electric vehicle charging load data time series, and decompose it into a plurality of electric vehicle charging load time series components;

[0009] S3: Calculate the fluctuation intensity of the original electric vehicle charging load time series and each electric vehicle charging load time series component based on sample entropy;

[0010] S4: According to the sample entropy value distribution of the original electric vehicle charging load time series and the electric vehicle charging load time series components, the electric vehicle charging load time series components are sequentially reconstructed into a trend component, a long time scale detail component, a short time scale detail component and a fluctuation component;

[0011] S5: Build an Attention-TCN neural network model;

[0012] S6: Construct the input variables of the Attention-TCN neural network;

[0013] S7: Calculate the output of the front-end attention module of the Attention-TCN neural network model;

[0014] S8: Calculate the output of the TCN module of the Attention-TCN neural network model;

[0015] S9: Calculate the output of the rear-end attention module of the Attention-TCN neural network model;

[0016] S10: Calculate the output of the Attention-TCN neural network model.

[0017] According to the above technical scheme, preferably, the original electric vehicle charging load data time series in S2 is u o (t), and the VMD algorithm is applied to decompose the original electric vehicle charging load data time series u o (t) into K electric vehicle charging load time series components, respectively represented as u1(t), u2(t), …, u K (t).

[0018] According to the above technical scheme, preferably, the calculation formula in S3 is:

[0019]

[0020] In equation (1) m v r is the pattern dimension, which is 1; v Similarity tolerance for each electric vehicle charging load component in the time series.

[0021] The original electric vehicle charging load time series u is obtained according to equation (1). o (t) and the time series of each electric vehicle charging load component u1(t), u2(t), ..., u K The sample entropy values ​​of (t) are s0, s1, s2, ..., s K .

[0022] According to the above technical solution, preferably, in S4:

[0023] Electric vehicle charging load time series components with sample entropy values ​​less than or equal to the original electric vehicle charging load time series sample entropy s0 are superimposed and reconstructed into trend components.

[0024] The electric vehicle charging load time series components whose sample entropy values ​​are greater than the original electric vehicle charging load time series sample entropy s0 and less than or equal to 2s0 are superimposed and reconstructed into long-term scale detail components.

[0025] Electric vehicle charging load time series components with sample entropy values ​​greater than 2s0 and less than or equal to 3s0 of the original electric vehicle charging load time series are superimposed and reconstructed into short time scale detail components.

[0026] The electric vehicle charging load time series components with sample entropy values ​​greater than the original electric vehicle charging load time series sample entropy 3s0 are superimposed and reconstructed into fluctuation components.

[0027] According to the above technical solution, preferably, S5 specifically includes,

[0028] S51: Construct the TCN model. The core of the TCN model is the residual dilated causal convolution unit. At the same time, in order to achieve input-output sequence alignment, m elements need to be added to the end of the input sequence of each residual dilated causal convolution unit, as shown in Equation (2).

[0029] m=(K-1)d (2)

[0030] In the formula, d is the expansion factor and K is the filter size;

[0031] S52: Construct the Attention-TCN model. Add an attention mechanism to the input and output of the TCN model.

[0032] According to the above technical solution, preferably, S6 specifically includes:

[0033] S61: input variables of an Attention-TCN neural network for trend component prediction are constructed,

[0034] The input is constructed by using a sliding window method, the window width is set to 96, that is, the historical values of the trend component at the previous 96 time points of the to-be-detected time point of the trend component are selected as the input of the model, the load values at the to-be-detected time point and the subsequent three time points are predicted, and the sliding window is moved forward by four time points each time.

[0035] S62: input variables of an Attention-TCN neural network for long-time-scale detail component prediction are constructed,

[0036] The input is constructed by using a sliding window method, the window width is set to 96, that is, the historical values of the long-time-scale detail component at the previous 96 time points of the to-be-detected time point of the long-time-scale detail component are selected as the input of the model, the load values at the to-be-detected time point and the subsequent three time points are predicted, and the sliding window is moved forward by four time points each time.

[0037] S63: input variables of an Attention-TCN neural network for short-time-scale detail component prediction are constructed,

[0038] The input is constructed by using a sliding window method, the window width is set to 96, that is, the historical values of the short-time-scale detail component at the previous 96 time points of the to-be-detected time point of the short-time-scale detail component are selected as the input of the model, the load values at the to-be-detected time point and the subsequent three time points are predicted, and the sliding window is moved forward by four time points each time.

[0039] S64: input variables of an Attention-TCN neural network for fluctuation component prediction are constructed,

[0040] The input is constructed by using a sliding window method, the window width is set to 16, that is, the historical values of the fluctuation component at the previous 16 time points of the to-be-detected time point of the fluctuation component are selected as the input of the model, the load values at the to-be-detected time point and the subsequent three time points are predicted, and the sliding window is moved forward by four time points each time.

[0041] According to the above technical scheme, preferably, S7 specifically comprises:

[0042] S71: the output vector c of the Attention-TCN at the t-1 time point is set to be t-1 , and the input quantity x is t-1 concatenated with the input quantity x to obtain [c t-1 ; x];

[0043] S72: the input quantity x is linearly transformed to obtain the query vector q of the front-end attention mechanism, and [c t-1; x] are respectively linearly changed to obtain the key vector k and the key value vector v of the front-end attention mechanism, as shown in formula (3),

[0044]

[0045] In the formula, W fq ,W fk ,W fv is a training parameter;

[0046] S73: Dot product operation is performed on the query vector q and each key vector k i to obtain the weight sim i between them, as shown in formula (4),

[0047]

[0048] S74: A softmax function is used to normalize the weight sim i , as shown in formula (5):

[0049]

[0050] S75: Finally, a i and the corresponding key value vector v i are weighted and summed to obtain the result, which is added to the input x to obtain the output r of the front-end attention mechanism, t as shown in formula (6),

[0051]

[0052] According to the above technical solution, preferably, S8 specifically includes:

[0053] r t is taken as the input of the TCN network, and the output after the TCN model is h t , as shown in formula (7),

[0054] h t = F(r t ) (7)

[0055] In the formula, F represents the TCN model.

[0056] According to the above technical solution, preferably, S9 specifically includes:

[0057] S91: The output vector h t of the t-th time TCN model and the output vectors [h t-L , h t-L+1 , …, h t-1 ] of the previous L time TCN models are spliced to form a set M tAs shown in formula (8),

[0058] M t = [h t-L ,h t-L+1 ,…,h t-1 ,h t ] (8);

[0059] S92: obtain the attention score vector β t at time t from the TCN model output vector h t and M t , as shown in formula (9).

[0060]

[0061] S93: obtain the output vector c t of the Attention-TCN at time t according to h t , β t and M t , as shown in formula (10),

[0062]

[0063] According to the above technical solution, preferably, S10 specifically comprises:

[0064] S101: repeat steps seven to nine to obtain the output vectors c t , c t+1 , c t+2 , c t+3 of the Attention-TCN at times t to t+3 of the day to be measured in turn;

[0065] S102: concatenate the output vectors c t , c t+1 , c t+2 , c t+3 of the Attention-TCN at times t to t+3 of the day to be measured into C, as shown in formula (11),

[0066] C=[c t ;c t+1 ;c t+2 ;c t+3 ] (11);

[0067] S103: after the output quantity C passes through the fully connected layer, obtain the load values y t , y t+1 , y t+2 , y t+3 at times t to t+3 of the day to be measured, as shown in formula (12),

[0068] yt ,y t+1 ,y t+2 ,y t+3 =f(C·W o1 +b o1 ) (12),

[0069] wherein f is a sigmod activation function; W o1 ,b o1 are training parameters.

[0070] The beneficial effects of the present application are: the prediction method of the present application applies variational mode decomposition (VMD) to decompose the electric vehicle charging load time series, and then recombines the sample entropy size of each component into a trend component, a long time scale detail component, a short time scale detail component, a fluctuation component and a residual component. Since VMD has good adaptive ability and can overcome the modal aliasing phenomenon, it can reflect the change law of the electric vehicle charging load time series at different time scales. In addition, due to the special residual dilation causal structure of the temporal convolutional network (TCN), it can still effectively capture the internal influence relationship of non-continuous data and long-interval historical data. On the basis of TCN, a double-end attention mechanism is added, which enables the model to adjust the output of the current prediction time step model by extracting the output information of the previous time step and the historical time step hidden information, and also adjusts the input of the TCN model according to the different prediction time steps, so that the model can dynamically update with time, thereby improving the accuracy of the electric vehicle charging load in the multi-step prediction case. BRIEF DESCRIPTION OF DRAWINGS

[0071] Figure 1 is a flowchart of the prediction method of the present application;

[0072] Figure 2 is the original electric vehicle charging load time series curve in the embodiment;

[0073] Figure 3 is the VMD decomposition result of the original electric vehicle charging load time series curve;

[0074] Figure 4 is the amplitude-frequency curve of each component after VMD decomposition of the original electric vehicle charging load time series;

[0075] Figure 5 is the amplitude-frequency curve of the original electric vehicle charging load time series;

[0076] Figure 6The sample entropy of the original electric vehicle charging load time sequence curve and each component after VMD decomposition thereof;

[0077] Figure 7 The original electric vehicle charging load trend quantity, long time scale detail quantity, short time scale detail quantity and fluctuation quantity time sequence curves obtained from the data shown in Figure 1

[0078] Figure 8 The structural schematic diagram of the Attention-TCN model;

[0079] Figure 9 The structural schematic diagram of the attention mechanism;

[0080] Figure 10 The curve diagram of the prediction result obtained by the prediction method and the actual electric vehicle charging load. DETAILED DESCRIPTION

[0081] In order to enable the person skilled in the art to better understand the technical solutions of the present application, the present application will be further described in detail below in combination with the drawings and the best embodiment.

[0082] Embodiment 1:

[0083] The historical electric vehicle charging load data of a city in North China from January 2018 to March 2020 is selected, the electric vehicle charging load data from January 2018 to February 2020 is selected as the training set, the measured electric vehicle charging load data of the first day in March 2020 is taken as the standard, the interval time between the electric vehicle charging load time points is 15 minutes, that is, the electric vehicle charging load of 96 time points needs to be predicted.

[0084] As shown in the figure, the present application provides an electric vehicle charging load prediction method based on an improved time convolution network, which comprises the following steps:

[0085] S1: Obtain the electric vehicle charging load data, and fill in the missing values, that is, fill in the missing values by using the data mean of the adjacent two time points;

[0086] S2: Apply the VMD algorithm to decompose the original electric vehicle charging load data time sequence u0(t) Figure 2 ) into 10 electric vehicle charging load time sequences u1(t), u2(t), …, u 10 (t) Figure 3 ), the amplitude-frequency curve of each component is shown in Figure 4 , and the amplitude-frequency curve of the original electric vehicle charging load is shown in Figure 5 .

[0087] ​S3: Calculate the sample entropy of each electric vehicle charging load component time series u k (t) is the complexity, i.e. fluctuation intensity, as shown in equation (1).

[0088]

[0089] m in equation (1) v is the pattern dimension, which is 1; r v is the similarity tolerance in each electric vehicle charging load time series component.

[0090] According to equation (1), the sample entropy value of each electric vehicle charging load time series component ( Figure 6 ) is obtained.

[0091] S4: According to the sample entropy value distribution of the electric vehicle charging load component time series, reconstruct the electric vehicle charging load components IMF1 and IMF2 into trend components, IMF3 and IMF4 into long time scale detail components, IMF5, IMF6 and IMF7 into short time scale detail components, and IMF8, IMF9 and IMF10 into fluctuation components ( Figure 7 ).

[0092] S5: Build an Attention-TCN neural network model ( Figure 8 )

[0093] S51: Build a TCN model. The core residual dilated causal convolution unit of the TCN model selects a dilation factor d = 1, 2, 4, and a filter size K = 3. At the same time, in order to ensure that the input and output sequences are aligned, (K-1) d elements need to be supplemented at the tail of the input sequence of each residual dilated causal convolution unit.

[0094] S52: Build an Attention-TCN model. Add an attention mechanism at the input and output ends of the TCN model ( Figure 9 ).

[0095] S6: Build the input variables of the Attention-TCN neural network.

[0096] S61: Build the input variables of the Attention-TCN neural network for trend component prediction.

[0097] Use the sliding window method to construct the input, set the window width to 96, i.e. select the trend component historical values of the previous 96 time points of the trend component to be measured time point as the input of the model, and predict the load values of the to-be-measured time point and the next 3 time points. The sliding window moves forward by 4 time points each time.

[0098] S62: Construct the input variable of the Attention-TCN neural network for long-time scale detail component prediction.

[0099] The input is constructed by using the sliding window method, and the window width is set to 96, that is, the historical values of the long-time scale detail component of the previous 96 time points of the to-be-measured time point of the trend component are selected as the input of the model to predict the load values of the to-be-measured time point and the next three time points. And the sliding window moves forward by 4 time points each time.

[0100] S63: Construct the input variable of the Attention-TCN neural network for short-time scale detail component prediction.

[0101] The input is constructed by using the sliding window method, and the window width is set to 96, that is, the historical values of the short-time scale detail component of the previous 96 time points of the to-be-measured time point of the trend component are selected as the input of the model to predict the load values of the to-be-measured time point and the next three time points. And the sliding window moves forward by 4 time points each time.

[0102] S64: Construct the input variable of the Attention-TCN neural network for fluctuation component prediction.

[0103] The input is constructed by using the sliding window method, and the window width is set to 16, that is, the historical values of the fluctuation component of the previous 16 time points of the to-be-measured time point of the trend component are selected as the input of the model to predict the load values of the to-be-measured time point and the next three time points. And the sliding window moves forward by 4 time points each time.

[0104] S7: Calculate the output of the front-end attention module of the Attention-TCN neural network model.

[0105] S71: Let the output vector of the Attention-TCN at time t-1 be c t-1 , and the input be x. Concatenate c t-1 and the input x to obtain [c t-1 ; x].

[0106] S72: After linear transformation, the input x obtains the query vector q of the front-end attention mechanism. After linear transformation, [c t-1 ; x] obtains the key vector k and the key value vector v of the front-end attention mechanism, as shown in formula (2).

[0107]

[0108] In the formula, W fq , W fk , and W fv are training parameters.

[0109] S73: Concatenate the query vector q and each key vector ki Do dot product operation, get the weight between the two sim i ; As shown in equation (3).

[0110]

[0111] S74: using a softmax function to weight sim i Normalization, as shown in equation (4).

[0112]

[0113] S75: finally, a i and the corresponding key value vector v i Weighted sum of the results in the input x, get the final front end attention mechanism output r t , as shown in equation (5).

[0114]

[0115] S8: calculate the output of Attention-TCN neural network model TCN module. r t As the input of TCN network, the output after the TCN model is h t , as shown in equation (6).

[0116] h t = F(r t ) (6)

[0117] In the formula, F represents the TCN model.

[0118] S9: calculate the output of the Attention-TCN neural network model backend attention module.

[0119] S91: by t time TCN model output vector h t And the output vector of the TCN model of the previous L time [h t-L , h t-L+1 , …, h t-1 ] Splicing constitutes a set M t , as shown in equation (7).

[0120] M t = [h t-L , h t-L+1 , …, h t …, ht-1 -1 , h t ] (7)

[0121] S92: by t time TCN model output vector h t And M t, the attention score vector β at time t is obtained t As shown in equation (8).

[0122]

[0123] S93: according to h t , β t and M t , the output vector c of Attention-TCN at time t is obtained t , as shown in equation (9).

[0124]

[0125] S10: calculate the output of the Attention-TCN neural network model.

[0126] S101: repeat steps seven to nine, and obtain the output vectors c t , c t+1 , c t+2 , c t+3 of Attention-TCN at time t to t+3 of the day to be tested in turn.

[0127] S102: splice the output vectors c t , c t+1 , c t+2 , c t+3 of Attention-TCN at time t to t+3 of the day to be tested into C, as shown in equation (10)

[0128] C=[c t ;c t+1 ;c t+2 ;c t+3 ] (10)

[0129] S103: after the output quantity C passes through the full connection layer, the load values y t , y t+1 , y t+2 , y t+3 at time t to t+3 of the day to be tested are obtained, as shown in equation (11).

[0130] y t ,y t+1 ,y t+2 ,y t+3 =f(C·W o1 +b o1 ) (11)

[0131] In the formula, f is the sigmod activation function; W o1 ,b o1 are training parameters.

[0132] According to the prediction method of the present application, the electric vehicle charging load at 96 time points on the first day of March 2020 is obtained.

[0133] The curve chart of the electric vehicle charging load at 96 time points on the first day of March 2020 obtained by the prediction method of the present application and the actually measured electric vehicle charging load is shown in Figure 10 The results show that the results obtained by the prediction method of the present application are close to the actually measured electric vehicle charging load, indicating that the prediction method of the present application has high accuracy.

[0134] The present application has the following beneficial effects: The prediction method of the present application applies variational mode decomposition (VMD) to decompose the electric vehicle charging load time series, and then recombines the sample entropy size of each component into a trend component, a long time scale detail component, a short time scale detail component, a fluctuation component and a residual component. Since VMD has good adaptive ability and can overcome the modal aliasing phenomenon, it can reflect the change law of the electric vehicle charging load time series at different time scales. In addition, due to the special residual dilated causal structure of the temporal convolutional network (TCN), it can still effectively capture the internal influence relationship of non-continuous data and long-interval historical data. On the basis of TCN, a double-end attention mechanism is added, which enables the model to adjust the output of the current prediction time step model by extracting the output information of the previous time step and the historical time step hidden information, and also adjusts the input of the TCN model according to the different prediction time steps, so that the model can dynamically update with time, thereby improving the accuracy of the electric vehicle charging load in the multi-step prediction case.

[0135] The above description is only the preferred embodiment of the present application, and it should be pointed out that for ordinary skilled persons in the technical field, some improvements and refinements can be made without departing from the principles of the present application, and these improvements and refinements should also be considered as the protection scope of the present application.

Claims

1. A method for predicting electric vehicle charging load based on an improved temporal convolutional network, characterized in that, Includes the following steps: S1: Obtain electric vehicle charging load data and fill in the missing values, that is, use the average of the data from two adjacent time points to fill in the missing values; S2: Use the time series of electric vehicle charging load data obtained in S1 as the original electric vehicle charging load data time series, and decompose it into multiple electric vehicle charging load time series components. S3: Calculate the original electric vehicle charging load time series and the fluctuation intensity of each electric vehicle charging load time series component based on sample entropy; S4: Based on the original electric vehicle charging load time series and the sample entropy distribution of the electric vehicle charging load time series components, the electric vehicle charging load time series components are reconstructed sequentially into trend components, long-term scale detail components, short-term scale detail components, and fluctuation components. S5: Building the Attention-TCN neural network model; details include, S51: Construct the TCN model; The core of the TCN model is the residual dilated causal convolution unit. At the same time, in order to achieve input-output sequence alignment, m elements need to be added to the end of the input sequence of each residual dilated causal convolution unit, as shown in Equation (2). m=(K-1)d (2) In the formula, d is the expansion factor and K is the filter size; S52: Construct an Attention-TCN model; add an attention mechanism to the input and output of the TCN model; S6: Input variables for constructing the Attention-TCN neural network; S7: Calculate the output of the front-end attention module of the Attention-TCN neural network model; S8: Calculate the output of the TCN module in the Attention-TCN neural network model; S9: Calculate the output of the backend attention module of the Attention-TCN neural network model; S10: Calculate the output of the Attention-TCN neural network model.

2. The electric vehicle charging load prediction method based on an improved temporal convolutional network according to claim 1, characterized in that, The original electric vehicle charging load data time series in S2 is u o (t), applying the VMD algorithm to transform the original electric vehicle charging load data time series u o The time series component (t) is decomposed into K electric vehicle charging load components, denoted as u1(t), u2(t), ..., u K (t).

3. The electric vehicle charging load prediction method based on an improved temporal convolutional network according to claim 2, characterized in that, The calculation formula in S3 is as follows: In equation (1) m v r is the pattern dimension, which is 1; v The similarity tolerance is set for each electric vehicle charging load component time series; the original electric vehicle charging load time series u is obtained according to equation (1). o (t) and the time series of each electric vehicle charging load component u1(t), u2(t), ..., u K The sample entropy values ​​of (t) are s0, s1, s2, ..., s K .

4. The electric vehicle charging load prediction method based on an improved temporal convolutional network according to claim 3, characterized in that, In S4: Electric vehicle charging load time series components with sample entropy values ​​less than or equal to the original electric vehicle charging load time series sample entropy s0 are superimposed and reconstructed into trend components. The electric vehicle charging load time series components whose sample entropy values ​​are greater than the original electric vehicle charging load time series sample entropy s0 and less than or equal to 2s0 are superimposed and reconstructed into long-term scale detail components. Electric vehicle charging load time series components with sample entropy values ​​greater than 2s0 and less than or equal to 3s0 of the original electric vehicle charging load time series are superimposed and reconstructed into short time scale detail components. The electric vehicle charging load time series components with sample entropy values ​​greater than the original electric vehicle charging load time series sample entropy 3s0 are superimposed and reconstructed into fluctuation components.

5. The electric vehicle charging load prediction method based on an improved temporal convolutional network according to claim 4, characterized in that, S6 specifically includes: S61: Construct the input variables for the Attention-TCN neural network for trend component prediction. Construct the input using the sliding window method. Set the window width to 96, that is, select the historical values ​​of the trend components at the 96th time before the time to be tested as the input of the model, and predict the load values ​​at the time to be tested and the next 3 times. The sliding window moves forward 4 times each time. S62: Construct the input variables for the Attention-TCN neural network for long-scale detail component prediction. Construct the input using the sliding window method. Set the window width to 96, that is, select the historical values ​​of the long-scale detail components of the time before the time to be tested as the input of the model, predict the load values ​​of the time to be tested and the next 3 times, and move the sliding window forward by 4 times each time. S63: Construct the input variables for the Attention-TCN neural network for short timescale detail component prediction. Construct the input using the sliding window method, set the window width to 96, that is, select the historical values ​​of the short timescale detail components of the previous 96 time steps of the time step to be tested as the input of the model, predict the load values ​​of the time step to be tested and the next 3 time steps, and move the sliding window forward by 4 time steps each time. S64: Construct the input variables for the Attention-TCN neural network for fluctuation component prediction. Construct the input using the sliding window method, set the window width to 16, that is, select the historical values ​​of the fluctuation component at the time before the time to be measured as the input of the model, predict the load value at the time to be measured and the next 3 times, and move the sliding window forward by 4 times each time.

6. The electric vehicle charging load prediction method based on an improved temporal convolutional network according to claim 5, characterized in that, Specifically, S7 includes: S71: Let the output vector of Attention-TCN at time t-1 be c. t-1 Given input x, set c t-1 Concatenating it with the input x yields [c t-1 ;x]; S72: The input x is linearly transformed to obtain the query vector q for the front-end attention mechanism, which is then used to [c t-1 ;x] respectively through linear After the transformation, the key vector k and key-value vector v of the front-end attention mechanism are obtained, as shown in equation (3). S73: Combine the query vector q and each key vector k i Perform a dot product operation to obtain the weight sim between the two. i As shown in equation (4), S74: Use a softmax function to simulate the weights. i Normalization is performed, as shown in equation (5): S75: Finally, a i and the corresponding key-value vector v i The weighted summation result is then added to the input x to obtain the final output r of the front-end attention mechanism. t As shown in equation (6), 。 7. The electric vehicle charging load prediction method based on an improved temporal convolutional network according to claim 6, Its features are, S8 specifically includes: r t As input to the TCN network, the output after passing through the TCN model is h. t As shown in equation (7), h t =F(r t ) (7) In the formula, F represents the TCN model.

8. The electric vehicle charging load prediction method based on an improved temporal convolutional network according to claim 7, characterized in that, S9 specifically includes: S91: The output vector h of the TCN model at time t t The output vector of the TCN model at the previous L time steps [h] t-L ,h t-L+1 ,…,h t-1 The sets M are constructed by splicing together. t As shown in equation (8), M t =[h t-L ,h t-L+1 ,…,h t-1 ,h t ] (8) S92: Output vector h from the TCN model at time t t and M t The attention score vector β at time t is obtained. t As shown in equation (9); S93: According to h t ,β t and M t Obtain the output vector c of Attention-TCN at time t. t As shown in equation (10), 。 9. The electric vehicle charging load prediction method based on an improved temporal convolutional network according to claim 8, characterized in that, S10 specifically includes: S101: Repeat S7 to S9 to obtain the output vector c of the Attention-TCN from time t to time t+3 on the day to be tested. t c t+1 c t+2 c t+3 ; S102: The output vector c of the Attention-TCN from time t to time t+3 on the day to be tested. t c t+1 c t+2 c t+3 The splicing is C, as shown in equation (11). C=[c t ;c t+1 ;c t+2 ;c t+3 ] (11) S103: After the output quantity C passes through the fully connected layer, the load value y from time t to time t+3 on the day under test is obtained. t y t+1 y t+2 y t+3 As shown in equation (12), y t ,y t+1 ,y t+2 ,y t+3 =f(C·W o1 +b o1 ) (12) In the formula, f is the sigmoid activation function; W o1 ,b o1 These are the training parameters.

Images