Electric vehicle cluster charging load interval prediction method based on deep learning

Abstract

The invention discloses an electric vehicle cluster charging load interval prediction method based on deep learning, and belongs to the technical field of electrical engineering. The method comprises the following steps: firstly, establishing a QR-LSTM neural network model; then, a data set containing the charging power and calendar information at the same historical time point is selected as an input variable, and on the basis, the calendar information is processed through a one-hot coding method to be converted into a numerical variable; thirdly, in the training process of the model, the data set is processed through a sliding window method, and normalization processing is carried out on the data; and the trained model can carry out interval prediction on the electric vehicle cluster charging load in the power distribution area. According to the method, the uncertainty of the charging load of the electric vehicle is considered, and the proposed prediction method can perform interval prediction on the charging load of the electric vehicle cluster; the prediction result can provide theoretical support for power grid planning, demand response and the like.

Description

Technical Field

[0001] The present invention relates to the field of electrical engineering technology, and in particular to a method for predicting charging load intervals of electric vehicle clusters based on deep learning. Background Art

[0002] In today's society, with the increasing awareness of environmental protection and the transformation of energy structure, electric vehicles are gradually becoming the mainstream choice as a clean energy means of transportation. However, the rapid popularization of electric vehicles has also brought a series of challenges, one of which is how to effectively manage and dispatch their charging needs to avoid adverse effects on the power grid. The charging demand of electric vehicles is concentrated and complex, which may lead to overload and instability of the power grid, and thus affect the normal life and work of residents. By predicting the charging load of electric vehicle clusters in distribution stations, possible load peaks can be predicted in advance, providing a scientific basis for power grid planning, ensuring that distribution system resources are efficiently configured, thereby improving power supply reliability and system resilience; at the same time, it can also provide theoretical support for the optimal scheduling, dynamic pricing and demand response strategies of electric vehicle charging. Therefore, it is urgent to study the prediction method of charging load of electric vehicle clusters in distribution stations.

[0003] There are two main types of existing electric vehicle cluster charging load prediction methods. One is to perform statistical analysis on the historical charging data of electric vehicles, establish models of variables such as charging duration, charging frequency, and energy consumption, and use these statistical models to generate typical predicted charging curves; the other is to use machine learning models such as RNN and LSTM to perform point prediction of electric vehicle charging load. However, both of the above methods have shortcomings. Electric vehicle charging load data has significant time series characteristics. The charging load is not only affected by the current moment, but also closely related to previous charging behaviors and trends. Traditional statistical methods often find it difficult to capture the complex dynamics of charging demand. At the same time, different users have different travel habits and charging habits. Even for the same electric vehicle user, their travel behavior and charging habits will be affected by subjective factors or multiple objective factors such as weather. Therefore, the charging load of the electric vehicle cluster will have a certain degree of uncertainty due to the user's charging habits. The traditional point prediction method will have a large error due to the uncertainty of the charging load. Unlike the point prediction method, the interval prediction method can provide prediction intervals at different confidence levels and effectively quantify the uncertainty of the charging load. Summary of the invention

[0004] In view of the problems existing in the prior art, the present invention proposes a method for predicting the charging load interval of an electric vehicle cluster based on deep learning.

[0005] To achieve the above object, the present invention adopts the following technical solution: a method for predicting charging load interval of electric vehicle cluster based on deep learning, comprising the following steps:

[0006] Step 1: Based on the LSTM neural network and combined with quantile regression, a QR-LSTM neural network model is constructed;

[0007] Step 2: The input variables of the QR-LSTM neural network model are the total charging power of electric vehicles in the distribution area at the same time point N days before the forecast and calendar information, and the output is the total charging power of electric vehicles in the distribution area, as shown below:

[0008] P A =Model(P T ,O);

[0009] Among them, P T =[P T1T ,P T2T ,...,P TNT ]∈R N×T To predict the total charging power of electric vehicles in the area at the same time N days ago, the length of the charging power sequence matrix is ​​T, P A ∈R 1×T To predict the total charging power sequence of electric vehicles in the Japan-Taiwan region with a length of T, O = [O 1T ,O 2T ...,O mT ]∈R N×T is a sequence matrix containing calendar information, and its length T is related to the interval of the prediction time points;

[0010] Step 3, using a data set containing historical charging data and calendar information of electric vehicles in the distribution station area to train the QR-LSTM neural network model; the training includes initializing weights, forward transmission, calculating losses, back propagation, and weight updating, and the training process is repeated until the performance of the QR-LSTM neural network model reaches a predetermined stop condition; using the trained QR-LSTM neural network model to predict the charging load of electric vehicles in the distribution station area.

[0011] Furthermore, the construction of the QR-LSTM neural network model is specifically as follows:

[0012] Establish an LSTM neural network, the basic unit of which defines three gates and an update node g t , the gate includes an input gate i t , output gate o t and forget gate f t ; In the basic unit of the LSTM neural network, the value of the cell state at time t is c t, is the core variable that transmits and stores information in long time series;

[0013] Assume that the input sequence of the basic unit of the LSTM neural network is {x1,x2,x3,...,x t}, where x t ∈R n Represents the value of an n-dimensional vector at time step t; the calculation form of each gate node of LSTM is as follows:

[0014] f t =σ(W f [x t ,h t-1 ]+b f )

[0015] i t =σ(W i [x t ,h t-1 ]+b i )

[0016] g t =σ(W g [x t ,h t-1 ]+b g )

[0017] o t =σ(W o [x t ,h t-1 ]+b o )

[0018] c t =g t ⊙i t +c t-1 ⊙f t

[0019] h t =tanh(c t )⊙o t ;

[0020] Among them, [,] means connecting the two vectors, w f 、w i 、w g 、w o Respectively represent the weight matrices of the forget gate, input gate, update node, and output gate, ⊙ represents element multiplication, σ represents the sigmoid activation function, and tanh represents the tanh activation function;

[0021] The definitions of the sigmoid function σ(z) and the tanh function tanh(z) are as follows:

[0022]

[0023] Among them, z represents the input variable, e is a constant, and represents the base of the natural logarithm function;

[0024] The loss function is modified to the loss function of quantile regression to obtain the QR-LSTM neural network model. The loss function of quantile regression is defined as:

[0025]

[0026] Among them, τ is the set quantile value, y is the true value, is the predicted value;

[0027] Set a set of digit values ​​τ = τ1, τ2, τ3, ..., τ N , N is the total number of quantile values; the QR-LSTM neural network model is trained N times, and the prediction intervals of electric vehicle charging load under different confidence levels are obtained by statistically analyzing the output N groups of results.

[0028] Furthermore, the calendar information includes season, day type and time information, and the day type includes weekdays, weekends and holidays; wherein the season and day type information are processed using a one-hot encoding method, specifically, 4 binary variables and 3 binary variables are used to process the season and day type information respectively, such as spring is represented by [1,0,0,0] and summer is represented by [0,1,0,0].

[0029] Furthermore, the training of the QR-LSTM neural network model includes: before training, using a sliding window method to process the data set containing the historical charging data and calendar information of the electric vehicles in the distribution station area; the sliding window length of the data sequence is m, the sliding window moves in steps of k each time, and the original data sequence is S = [s1, s1, ..., s n ], then the i-th data sequence sample S obtained by the sliding window method i for:

[0030] S i =[s (k-1)*i+1 ,s (k-1)*i+2 ,...,s (k-1)*i+m ].

[0031] The sliding window method uses windows in time series data and reuses data during the training process, which helps increase the effectiveness of training data, improves training efficiency, and makes better use of limited data resources.

[0032] Furthermore, the training of the QR-LSTM neural network model using a data set containing historical charging data and calendar information of electric vehicles in the distribution station area includes: normalizing the data because the electric vehicle charging power and calendar information contained in the input vector have different dimensions; using the Z-score normalization method to standardize each dimension of the feature, and the processed data conforms to a standard normal distribution with a mean of 0 and a variance of 1; for the i-th sample x of the k-th dimension feature ki , after standardization, we get As shown below:

[0033]

[0034] Among them, μ k is the average value of the original data of the k-th dimension feature, σ k is the standard deviation of the original data of the k-th dimension feature.

[0035] Compared with the prior art, the present invention has the following beneficial effects:

[0036] 1. Taking into account the uncertainty factors of electric vehicle charging load, the present invention introduces quantile regression on the basis of LSTM neural network. The central tendency and statistical dispersion provided by quantile regression provide a more comprehensive data distribution perspective, so that the QR-LSTM model can output the prediction interval of electric vehicle cluster charging load under different confidence levels.

[0037] 2. The model proposed in this invention comprehensively considers various factors that affect the charging load of electric vehicles, including historical charging power, calendar information such as seasons and holidays.

[0038] 3. In the process of training the model, the present invention uses the sliding window method to process the data set, which helps to increase the effectiveness of the training data and make better use of limited data resources. At the same time, for time series data such as electric vehicle charging load, the sliding window method helps the neural network model capture the short-term and long-term dependencies of the data.

[0039] Considering that the current point prediction method for electric vehicle charging load prediction is difficult to solve the uncertainty problem of electric vehicle charging load, the present invention proposes a method for predicting the interval of electric vehicle cluster charging load based on deep learning, introduces quantile regression on the basis of LSTM neural network, and establishes a QR-LSTM model. The model fully considers the influence of factors such as calendar information on electric vehicle charging load, and can output the prediction interval of electric vehicle cluster charging load under different confidence levels. The prediction results can provide theoretical support for power grid planning, optimal scheduling of electric vehicle charging, dynamic pricing and demand response strategy. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 This is a framework diagram of the electric vehicle cluster charging load interval prediction method of the present invention;

[0041] Figure 2 Schematic diagram of LSTM neurons;

[0042] Figure 3 Schematic diagram of the sliding window method. DETAILED DESCRIPTION

[0043] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0044] like Figure 1 As shown, the embodiment of the present invention proposes a method for predicting charging load intervals of electric vehicle clusters based on deep learning, comprising the following steps:

[0045] First, a QR-LSTM neural network model is constructed. The QR-LSTM neural network model combines quantile regression with LSTM neural network, so that the LSTM neural network can perform interval prediction. The QR-LSTM neural network model structure of the present invention consists of an input layer, an LSTM layer, a fully connected layer, a quantile layer and an output layer. The input layer takes the charging power sequence and calendar information of the same time point in the history of the electric vehicle as input, the LSTM layer is used to process the sequence data of the input layer, the fully connected layer can enable the network to better extract features and learn input data, and the quantile layer is used to process multiple outputs of different quantiles and output by the output layer.

[0046] The LSTM layer is composed of LSTM basic units. The structure of LSTM basic units is as follows: Figure 2 As shown, three gates are defined in the LSTM basic unit (input gate i t , output gate o t and forget gate f t ) and update node g t In the basic unit of the LSTM neural network, the cell state is the core variable that transmits and stores information in a long time series, and its value at time t is c t ,The cell state plays an important role in LSTM neural networks learning the dependencies existing in long time series data.

[0047] Assume that the input sequence of the LSTM basic unit is {x1,x2,x3,...,x t}, where x t ∈R n , represents the value of the n-dimensional vector at time step t. The calculation form of each LSTM gate node is as follows:

[0048] f t =σ(W f [xt ,h t-1 ]+b f )

[0049] i t =σ(W i [x t ,h t-1 ]+b i )

[0050] g t =σ(W g [x t ,h t-1 ]+b g )

[0051] o t =σ(W o [x t ,h t-1 ]+b o )

[0052] c t =g t ⊙i t +c t-1 ⊙f t

[0053] h t =tanh9c t )⊙o t ;

[0054] In the formula, [,] means connecting two vectors, wf and w i 、w g 、w o are the weight matrices of the forget gate, input gate, update node, and output gate, respectively. ⊙ represents element multiplication, σ represents the sigmoid activation function, and tanh represents the tanh activation function. The definitions of the sigmoid function σ(z) and the tanh function tanh(z) are as follows:

[0055]

[0056] Among them, σ(z) represents the sigmoid activation function, tanh(z) represents the tanh activation function, z represents the input variable, and e is a constant representing the base of the natural logarithm function.

[0057] The loss function is modified to the loss function of quantile regression to obtain the QR-LSTM neural network model. The loss function of quantile regression can be defined as:

[0058]

[0059] Among them, τ is the set quantile value, y is the true value, is the predicted value.

[0060] Set a set of digit values ​​τ = τ1, τ2, τ3, ..., τ N , N is the total number of quantile values; the QR-LSTM neural network model is trained N times, and the prediction intervals of electric vehicle charging load under different confidence levels are obtained by statistically analyzing the output N groups of results.

[0061] Then, the input variables of the QR-LSTM neural network are set to predict the total charging power of electric vehicles in the distribution area at the same time N days ago and the calendar information, and the output is the total charging power of electric vehicles in the distribution area, as shown below:

[0062] P A =Model(P T ,O);

[0063] Among them, P T =[P T1T ,P T2T ,...,P TNT ]∈R N×T To predict the total charging power of electric vehicles in the area at the same time N days ago, the length of the charging power sequence matrix is ​​T, P A ∈R 1×T To predict the total charging power sequence of electric vehicles in the Japan-Taiwan region with a length of T, O = [O 1T ,O 2T ...,O mT ]∈R N×T is a sequence matrix containing calendar information, and its length T is related to the interval of the predicted time points.

[0064] Considering that the input features of the QR-LSTM neural network are the total charging power of electric vehicles in the area at the same time point N days before the prediction date and calendar information, and the calendar information includes season, day type (weekday / weekend / holiday) and time information, in order to use the calendar information as input features for training, the calendar information needs to be converted into numerical variables. Therefore, the one-hot encoding method is used to process the season and day type information, that is, 4 binary variables and 3 binary variables are used for processing respectively, such as [1,0,0,0] for spring and [0,1,0,0] for summer.

[0065] Then, after the QR-LSTM model is established, the QR-LSTM neural network model is trained using a data set containing historical charging data and calendar information of electric vehicles in the distribution station area. The training steps include initializing weights, forward pass, calculating losses, back propagation, and weight update. The training process is repeated until the network performance reaches a satisfactory level or reaches a predetermined stop condition.

[0066] Before training the QR-LSTM neural network model, in order to improve the training efficiency and make full use of limited data resources, the data set is first processed using the sliding window method before training. The sliding window method can effectively capture the time series information in the data, learn the time correlation in the data, and help the neural network model capture the short-term and long-term dependencies of the data. The schematic diagram of the sliding window method is shown in Figure 3 As shown, the sliding window length of the data sequence is m, the sliding window moves in steps of k each time, and the original data sequence is S = [s1, s1, ..., s n ], then the i-th data sequence sample S obtained by the sliding window method i for:

[0067] S i =[s (k-1)*i+1 ,s (k -1)*i+2,...,s (k-1)*i+m ];

[0068] When inputting historical data into the neural network for training, the data needs to be normalized because the electric vehicle charging power and calendar information contained in the input vector have different dimensions. The present invention uses the Z-score normalization method to normalize each dimension of the feature, and the processed data conforms to the standard normal distribution with a mean of 0 and a variance of 1. ki , after standardization, we get As shown below:

[0069]

[0070] Among them, μ k is the average value of the original data of the k-th dimension feature, σ k is the standard deviation of the original data of the k-th dimension feature.

[0071] After setting appropriate hyperparameters and structural parameters (number of hidden layers, number of neurons in each layer, learning rate, initial weight) of the QR-LSTM neural network model, the QR-LSTM neural network model is trained. The trained QR-LSTM neural network model is used to predict the charging load of the electric vehicle cluster in the distribution station area, and the prediction results are multiple prediction intervals with different confidence levels.

[0072] Those skilled in the art will readily appreciate other embodiments of the present application after considering the description and practicing the contents disclosed herein. The present application is intended to cover any variations, uses or adaptations of the present application, which follow the general principles of the present application and include common knowledge or customary technical means in the art that are not disclosed in the present application. The description and examples are intended to be exemplary only.

[0073] It should be understood that the present application is not limited to the exact construction that has been described above and shown in the drawings, and that various modifications and changes may be made without departing from the scope thereof.

Claims

1. A method for predicting charging load intervals of electric vehicle clusters based on deep learning, characterized in that: The following steps are involved: Based on the LSTM neural network and combined with quantile regression, a QR-LSTM neural network model is constructed; The input variables of the QR-LSTM neural network model are the total charging power of electric vehicles in the distribution area at the same time point N days before the forecast and calendar information, and the output is the total charging power of electric vehicles in the distribution area, as shown below: P A =Model(P T ,ABOUT); Among them, P T =[P T1T ,P T2T ,...,P TNT ]∈R N×T To predict the total charging power of electric vehicles in the area at the same time N days ago, the length of the charging power sequence matrix is ​​T, P A ∈R 1×T To predict the total charging power sequence of electric vehicles in the Japan-Taiwan region with a length of T, O = [O 1T ,O 2T ...,O mT ]∈R N×T is a sequence matrix containing calendar information, and its length T is related to the interval of the prediction time points; A QR-LSTM neural network model is trained using a data set containing historical charging data and calendar information of electric vehicles in a distribution station area; the training includes initializing weights, forward propagation, calculating losses, back propagation, and weight updating, and the training process is repeated until the performance of the QR-LSTM neural network model reaches a preset stop condition; the trained QR-LSTM neural network model is used to predict the charging load of electric vehicles in the distribution station area.

2. The method for predicting charging load interval of electric vehicle cluster based on deep learning according to claim 1 is characterized in that: The construction of the QR-LSTM neural network model is specifically as follows: Establish an LSTM neural network, the basic unit of which defines three gates and an update node g t , the gate includes an input gate i t , output gate o t and forget gate f t ; In the basic unit of the LSTM neural network, the value of the cell state at time t is c t , is the core variable that transmits and stores information in long time series; Assume that the input sequence of the basic unit of the LSTM neural network is {x1,x2,x3,...,x t }, where x t ∈R n Represents the value of an n-dimensional vector at time step t; the calculation form of each gate node of LSTM is as follows: f t =σ(W f [x t ,h t-1 ]+b f ) i t =σ(W i [x t ,h t-1 ]+b i ) g t =σ(W g [x t ,h t-1 ]+b g ) the t =σ(W o [x] t ,h t-1 ]+b o ) c t =g t ⊙i t +c t-1 ⊙f t h t =tanh(c t )⊙o t ; Among them, w f 、w i 、w g 、w o Respectively represent the weight matrices of the forget gate, input gate, update node, and output gate, ⊙ represents element multiplication, σ represents the sigmoid activation function, and tanh represents the tanh activation function; The definitions of the sigmoid function σ(z) and the tanh function tanh(z) are as follows: Among them, z represents the input variable, e is a constant, and represents the base of the natural logarithm function; The loss function is modified to the loss function of quantile regression to obtain the QR-LSTM neural network model. The loss function of quantile regression is defined as: Among them, τ is the set quantile value, y is the true value, is the predicted value; Set a set of digit values ​​τ = τ1, τ2, τ3, ..., τ N , N is the total number of quantile values; the QR-LSTM neural network model is trained N times, and the prediction intervals of electric vehicle charging load under different confidence levels are obtained by statistically analyzing the output N groups of results.

3. The method for predicting charging load interval of electric vehicle cluster based on deep learning according to claim 1 is characterized in that: The calendar information includes season, day type and time information, and the day type includes weekdays, weekends and holidays; wherein the season and day type information are processed using a one-hot encoding method, specifically, 4 binary variables and 3 binary variables are used to process the season and day type information respectively.

4. The method for predicting charging load interval of electric vehicle clusters based on deep learning according to claim 1 is characterized in that: The training of the QR-LSTM neural network model includes: before training, using a sliding window method to process the data set containing the historical charging data and calendar information of the electric vehicles in the distribution station area; the sliding window length of the data sequence is m, the sliding window moves in steps of k each time, and the original data sequence is S=[s1,s1,...,s n ], then the i-th data sequence sample S obtained by the sliding window method i for: S i =[s (k-1)*i+1 ,s (k-1)*i+2 ,...,s (k-1)*i+m ]。 5. The method for predicting charging load interval of electric vehicle cluster based on deep learning according to claim 1 is characterized in that: The method of training the QR-LSTM neural network model using a data set containing historical charging data and calendar information of electric vehicles in a distribution station area includes: normalizing the data; using a Z-score normalization method to normalize each dimension of the feature, so that the processed data conforms to a standard normal distribution with a mean value of 0 and a variance of 1; for the i-th sample x of the k-th dimension feature ki , after standardization, we get As shown below: Among them, μ k is the average value of the original data of the k-th dimension feature, σ k is the standard deviation of the original data of the k-th dimension feature.

Images