Charging station utilization prediction method and system based on step-by-step decomposition model

By combining the step-by-step decomposition model with historical and weather data to predict the utilization rate of charging stations, the problem of difficult to predict the utilization rate of charging stations in the existing technology is solved, and efficient positioning of idle charging piles is achieved, which improves user experience and grid stability.

CN116644853BActive Publication Date: 2024-11-19STATE GRID JIANGSU ELECTRIC POWER CO LTD NANJING POWER SUPPLY COMPANY
View PDF 4 Cites 0 Cited by

Patent Information

Application Number
CN202310617851.X
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2023-05-29
Publication Date
2024-11-19
Estimated Expiration
2043-05-29

AI Technical Summary

Technical Problem

There are few researches on the utilization rate forecast of existing charging stations, which is difficult to help users find charging stations containing idle charging piles efficiently, resulting in electric car owners encountering traffic congestion and other problems when looking for charging.

Method used

The charging station utilization prediction method based on the step-by-step decomposition model is adopted, and the utilization rate of future charging stations is predicted through the deep information fusion layer and the residual decomposition prediction layer, combined with historical utilization data, weather data and date data. This model uses attention coding method to fusion information, and refines prediction step by step through residual decomposition to improve prediction accuracy.

Benefits of technology

It realizes high-precision prediction of charging station utilization, helps users to find idle charging piles efficiently, saves time and costs, improves travel experience, and improves the stability of the power grid.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN116644853B_ABST
    Figure CN116644853B_ABST
Patent Text Reader

Abstract

The present invention relates to a charging station utilization prediction method based on a step-by-step decomposition model, comprising: collecting historical charging station utilization data, as well as corresponding weather data and date data, as raw data; constructing and training a step-by-step decomposition prediction model, inputting the raw data before the time period to be tested into the step-by-step decomposition prediction model, and obtaining a predicted value of the charging station utilization; the step-by-step decomposition prediction model comprises a deep information fusion layer and a residual decomposition prediction layer; the deep information fusion layer uses an attention encoding method to perform deep extraction and fusion of input data, obtains an information fusion sequence and inputs it into the residual decomposition prediction layer; the residual decomposition prediction layer uses a residual decomposition method to decompose the current layer input data step by step, and obtains prediction sequences and residual sequences of different fine-grainedness at each level, and finally adds the prediction sequences of all levels and the residual sequence of the last layer through the fully connected layer to obtain the final prediction result. The present invention realizes the prediction of charging station utilization.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of electric power systems, and in particular to a method and system for predicting charging station utilization based on a step-by-step decomposition model. Background Art

[0002] With the rapid development of society, the new power system Internet of Things with new energy as the main body has gradually become an important part of the power system. Among them, the charging station is the key hub connecting the grid side and the user side in the power Internet of Things. However, with the continuous acceleration of energy transformation, the number of electric vehicles has increased rapidly, and problems such as insufficient charging piles, difficulty in finding available charging piles, and traffic congestion caused by finding idle charging piles have become increasingly prominent. In the face of the above problems, blindly building new charging stations is not an economical and sustainable solution. By predicting the utilization rate of each charging station, users can be effectively guided to the charging station with idle charging piles, thereby helping electric car owners solve charging problems economically and efficiently.

[0003] The utilization prediction of charging stations can be regarded as a time series prediction problem. Traditional time series prediction techniques are mainly regression analysis prediction methods, such as autoregressive model (AR), autoregressive moving average model (ARMA), etc. Artificial neural network models have been used by many researchers in the field of time series prediction in recent years because of their learning ability to approximate any nonlinear function. Representative models include: long short-term memory recurrent neural network (LSTM) and gated recurrent unit structure (GRU). For existing research on charging station time series prediction, most researchers focus on the load prediction of charging stations or charging piles, so as to help power management departments better deal with load fluctuations when charging stations are connected to the power grid and improve the stability of the power grid. However, few researchers have predicted the utilization of charging stations to help users more efficiently find charging stations with idle charging piles and improve their travel experience.

[0004] In view of this, in order to improve the above-mentioned deficiencies, providing a charging station utilization prediction method and system based on a step-by-step decomposition model has become an urgent problem to be solved in this field. Summary of the invention

[0005] The purpose of the present invention is to provide a charging station utilization prediction method based on a step-by-step decomposition model to achieve the prediction of charging station utilization with high prediction accuracy.

[0006] In order to solve the above technical problems, the technical solution of the present invention is:

[0007] In a first aspect, a method for predicting charging station utilization based on a level-by-level decomposition model is provided, comprising:

[0008] Step S100: collecting historical charging station utilization data, as well as corresponding weather data and date data, as raw data; Step S200: preprocessing the raw data;

[0009] Step S300: constructing a step-by-step decomposition prediction model for charging station utilization prediction and training the step-by-step decomposition prediction model based on original data, wherein the step-by-step decomposition prediction model includes: a deep information fusion layer and a residual decomposition prediction layer;

[0010] The original data is input into the deep information fusion layer as input data. The deep information fusion layer uses the attention encoding method to deeply extract and fuse the coupling information in the input data, obtains the information fusion sequence and inputs it into the residual decomposition prediction layer;

[0011] The residual decomposition prediction layer uses the residual decomposition method to decompose the input data of the current layer step by step, and obtains prediction sequences and residual sequences of different fine-grained levels at each level. Finally, the prediction sequences of all levels and the residual sequence of the last layer through the fully connected layer are added to obtain the final prediction result;

[0012] Step S400: inputting historical charging station utilization data, weather data and date data before the time period to be tested into the trained step-by-step decomposition prediction model to obtain a predicted value of charging station utilization within a preset time in the future.

[0013] Specifically, the historical charging station utilization data, weather data, and date data at T time points before the prediction period are input into the step-by-step decomposition prediction model, and a prediction sequence with a length of H is output, i.e., the predicted value of the charging station utilization at H time points in the future.

[0014] Further, in step S100, the weather data selects two factors that have a greater impact on travel, namely, temperature data and visibility data;

[0015] Date data is accurate to the month and day, and a holiday label is added after it. The month and day are encoded with 12 and 31 digits respectively, and the holiday label is encoded with the corresponding number of digits according to the number of holiday types. Assuming there are 10 holidays in total, the number of digits of the holiday unique hot encoding is 10.

[0016] Concatenate the unique hot codes of month, date, and holiday in sequence and add 0 after them so that the length of the date data is consistent with the length of the time series subsequently fed into the model.

[0017] Further, in step S200, data preprocessing includes data cleaning and normalization operations;

[0018] Data cleaning includes correcting outliers in utilization data and weather data in the original data and filling in missing values. After completing the correction of outliers and filling in missing values ​​in utilization data and weather data, normalization operations are performed on each type of data.

[0019] Furthermore, in step S300, the working method of the deep information fusion layer is: taking the historical charging station utilization data of T time points before the prediction period, as well as the corresponding weather data and date data, assuming that there are N types of features in total, that is, decomposing the input sequence of the prediction model level by level into a sequence with a length of T and a dimension of N; sending the input sequence to the deep information fusion layer to obtain an information fusion sequence; specifically including:

[0020] First, the deep information fusion layer inserts a randomly initialized marker sequence based on the standard normal distribution at the first position of the input sequence;

[0021] Secondly, the deep information fusion layer uses the attention encoding method for information fusion, specifically: first, the input sequence containing the mark sequence with a length of T and a dimension of N+1 is multiplied by three different parameter matrices of size (N+1)*(N+1) to obtain the query matrix Q, the key matrix K and the value matrix V respectively, then the query matrix Q and the key matrix K are multiplied to obtain the attention matrix, and then the attention matrix is ​​normalized by the softmax function, and then the normalized attention matrix is ​​multiplied by the value matrix V to obtain the output sequence of the attention encoding method; the attention encoding method can learn the correlation between different sequences through sufficient data interaction between input sequences, and the introduction of the mark sequence enables it to obtain the universal correlation relationship between the original input sequences, that is, the information fusion of the input data is completed;

[0022] Finally, the deep information fusion layer extracts the subsequence at the first position of the attention encoding method output sequence and sends it to the fully connected layer to further extract deep information, and obtains the information fusion sequence of length T output by the deep information fusion layer. The information fusion sequence is then sent to the residual decomposition prediction layer to obtain the final prediction result, and the time length of the prediction result is set to H.

[0023] Furthermore, in step S300, the working method of the residual decomposition prediction layer is:

[0024] The residual decomposition prediction layer consists of multiple hierarchical components. The output of each hierarchical component includes two parts: the hierarchical prediction sequence and the inter-layer residual sequence. The length of the hierarchical prediction sequence is H, and the length of the inter-layer residual sequence is T.

[0025] The input of each hierarchical component is the hierarchical residual sequence output by the previous hierarchical component (the input of hierarchical component 1 is the information fusion sequence). Finally, the hierarchical prediction sequence output by all hierarchical components and the inter-layer residual sequence reconstructed by the fully connected layer output by the last layer are added together to obtain the final prediction result.

[0026] The hierarchical component consists of multiple decomposition modules. The output of each decomposition module includes two parts: module prediction sequence and module backtracking sequence. The length of the module prediction sequence is H, and the length of the module backtracking sequence is T.

[0027] The input of each decomposition module is the residual of the input sequence of the previous decomposition module and the module traceback sequence output by the previous decomposition module (the input of decomposition module 1 is the corresponding hierarchical input sequence). Finally, the module prediction sequences output by all decomposition modules are added together to obtain the hierarchical prediction sequence output by the hierarchical component. The residual of the input of the last decomposition module and the module traceback sequence output by the decomposition module is the inter-layer residual sequence output by the hierarchical component.

[0028] The decomposition module is the basic unit of the residual decomposition prediction layer, which consists of one-dimensional convolution, multi-layer perceptron and interpolation function. The input of the decomposition module is first downsampled by one-dimensional convolution, and then the downsampled sequence is passed through two different multi-layer perceptrons to reconstruct it into two sets of parameter sequences θ for subsequent interpolation operations. f and θ b ; Finally, the above two sets of parameter sequences are brought into the interpolation function, and the two sets of parameter sequences are respectively extended to the prediction length H of the model output and the length T of the model input, that is, the module prediction sequence and the module backtracking sequence are obtained.

[0029] Furthermore, in the decomposition module, the downsampling operation sets different sampling multiples according to the different positions of the hierarchical components. Specifically, the more forward the hierarchical component is, the greater the downsampling multiple of the decomposition module therein is.

[0030] The implementation method of downsampling is: downsampling is achieved by controlling the step size of the convolution kernel in the one-dimensional convolution. The larger the step size, the greater the downsampling multiple.

[0031] Furthermore, in the decomposition module, the interpolation function adopts nearest neighbor interpolation, linear interpolation or polynomial interpolation.

[0032] Furthermore, in step S300, the mean square error is used as the loss function of the step-by-step decomposition prediction model. Through the gradient descent algorithm, the parameters in the model will be continuously iteratively updated, so that the loss function value will continue to decrease. When the loss function value hardly changes, it means that the prediction model has converged, the training process is completed, and the parameters of the model are fixed.

[0033] In a second aspect, the present invention also provides a charging station utilization prediction system based on a hierarchical decomposition model, comprising a memory and a processor; wherein the memory stores a computer program, and when the program is executed by the processor, it can implement the above-mentioned charging station utilization prediction method based on the hierarchical decomposition model.

[0034] In a third aspect, the present invention further provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the above-mentioned charging station utilization prediction method based on the step-by-step decomposition model.

[0035] The present invention has the following beneficial effects:

[0036] 1. The present invention proposes a method for predicting the utilization rate of charging stations. From the user's perspective, the method uses historical charging station utilization data, weather data and date data to predict the utilization rate of charging stations in the future through a constructed hierarchical decomposition model. Users can efficiently find charging stations with idle charging piles based on the predicted utilization rate of charging stations, thereby saving the user's time cost and improving the user's travel experience.

[0037] 2. The present invention proposes a step-by-step decomposition prediction model. The network structure of the proposed step-by-step decomposition prediction model mainly includes: a deep information fusion layer and a residual decomposition prediction layer. Among them, the deep information fusion layer uses the attention encoding method of introducing the mark sequence to realize the deep extraction and fusion of coupled information in multi-category input data, and outputs the information fusion sequence. The information fusion sequence characterizes the typical characteristics of the input data, so it can accelerate the convergence of the subsequent prediction layer and make the prediction results more accurate. The architecture of the residual decomposition prediction layer realizes two functions. First, by controlling the different downsampling multiples in the decomposition module, the receptive field in the module input sequence is gradually refined, that is, the granularity of the prediction sequence is continuously improved; second, by repeatedly performing residual re-decomposition between the decomposition modules, the progressive learning composed of multiple hierarchical components is completed, that is, the subsequent decomposition module learns the content that the previous decomposition module has not learned. From the above two points, the residual decomposition prediction layer realizes the improvement of the overall prediction performance of the step-by-step decomposition model. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 is a flow chart of the prediction method of the present invention;

[0039] Figure 2 A schematic diagram of a level-by-level decomposition prediction model structure in an embodiment of the present invention;

[0040] Figure 3 This is a schematic diagram of a deep information fusion layer in an embodiment of the present invention;

[0041] Figure 4Schematic diagram of residual decomposition prediction layer in an embodiment of the present invention. DETAILED DESCRIPTION

[0042] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0043] Please refer to Figure 1 The present invention is a method for predicting charging station utilization based on a step-by-step decomposition model, which includes:

[0044] Step S100: collecting historical charging station utilization data, as well as corresponding weather data and date data, as raw data; Step S200: preprocessing the raw data;

[0045] Step S300: constructing a step-by-step decomposition prediction model for charging station utilization prediction and training the step-by-step decomposition prediction model based on original data, wherein the step-by-step decomposition prediction model includes: a deep information fusion layer and a residual decomposition prediction layer;

[0046] The original data is input into the deep information fusion layer as input data. The deep information fusion layer uses the attention encoding method to deeply extract and fuse the coupling information in the input data, obtains the information fusion sequence and inputs it into the residual decomposition prediction layer;

[0047] The residual decomposition prediction layer uses the residual decomposition method to decompose the input data of the current layer step by step, and obtains prediction sequences and residual sequences of different fine-grained levels at each level. Finally, the prediction sequences of all levels and the residual sequence of the last layer through the fully connected layer are added to obtain the final prediction result;

[0048] Step S400: inputting historical charging station utilization data, weather data and date data before the time period to be tested into the trained step-by-step decomposition prediction model to obtain a predicted value of charging station utilization within a preset time in the future.

[0049] The following are Figure 1 Each step in is described in detail.

[0050] In step S100, historical charging station utilization data, as well as corresponding weather data and date data are collected as raw data;

[0051] In this embodiment, weather data and date data constitute the influencing factor data. Among them, the weather data selects two factors that have a greater impact on travel, namely temperature data and visibility data. The date data is accurate to the month and date, and a holiday label is added after it. The month and date are uniquely hot encoded with 12 and 31 bits respectively, and the holiday label is uniquely hot encoded with the corresponding number of bits according to the number of holiday types. Assuming that there are 10 kinds of holidays, the number of bits of the holiday unique hot encoding is 10. The unique hot codes of the month, date and holiday are concatenated together in sequence, and 0 is added after them, so that the length of the date data is consistent with the length of the time series subsequently fed into the model. The prediction goal is to use historical data to predict the utilization rate of the charging station at the next 24 time points. The time interval between each two time points is 15 minutes, that is, to predict the utilization rate of the charging station in the next 3 hours.

[0052] In step S200, the original data is preprocessed;

[0053] In this embodiment, data preprocessing includes data cleaning and normalization operations. It should be noted that the date data in the original data is only a feature label used to help the prediction model capture the long-term characteristics and unconventional time characteristics of the utilization time series, so there is no need to perform data cleaning and normalization operations on it.

[0054] Data cleaning includes the correction of outliers in utilization data and weather data in the original data and the filling of missing values. The method for judging outliers in utilization data and weather data is the 3sigma criterion, and its formula is as follows:

[0055]

[0056] Among them, x n,t represents the value at time t on day n, represents the mean of all moments on the nth day, and σ represents the standard deviation of all moments on the nth day. For points that meet the above formula, they are judged as outlier points, and the mean of the date they are located is taken as the correction value for the outlier points. The formula for correcting outliers is as follows:

[0057]

[0058] Among them, x n,t ′ is the corrected value at time t on day n, Represents the mean of all sampling times on day n.

[0059] For missing values ​​in utilization data and weather data, if the number of missing value points is greater than or equal to one-third of the number of record points in a day, the data of that day will be discarded; if the number of missing points is less than one-third of the number of record points in a day, the missing values ​​will be filled based on the data of the two moments before and after the same day and the data of the same moment two days before and after. The formula for filling missing values ​​is as follows:

[0060]

[0061] Among them, x n,t ′ is the value at time t on the nth day after filling, x n+1,t and x n-1,t are the values ​​at time t on day n+1 and day n-1, respectively, n,t+1 and x n,t-1 are the values ​​at time t+1 and time t-1 on day n, respectively. α1 and α2 are weight coefficients, and their sum should be equal to 1.

[0062] After completing the correction of outliers and filling of missing values ​​in utilization data and weather data, each type of data needs to be normalized to eliminate the impact of dimensions between different types of data and improve the accuracy of the prediction model. The normalization method uses the min-max normalization method. For any set of data, the normalization formula is as follows:

[0063]

[0064] in, represents the normalized value at time t on day n, d n,t Indicates the value at time t on day n after data cleaning, maxd n and mind n They represent the maximum and minimum values ​​at all times on the nth day after data cleaning.

[0065] In step S300, a level-by-level decomposition prediction model is constructed and trained.

[0066] The network structure of the prediction model is decomposed step by step. Figure 2As shown in the figure, it mainly includes: deep information fusion layer and residual decomposition prediction layer. Among them, the deep information fusion layer uses the attention encoding method to realize the deep extraction and fusion of coupled information in multi-category input data, and outputs the information fusion sequence. The information fusion sequence represents the typical characteristics of the input data, so it can accelerate the convergence of the subsequent prediction layer and make the prediction results more accurate. The residual decomposition prediction layer uses the residual decomposition method to decompose the input data of this layer step by step. Each level obtains prediction sequences and residual sequences of different fine-grained levels through the decomposition module. Finally, the prediction sequences of all levels and the residual sequence of the last layer through the fully connected layer are added to obtain the final prediction result. This design mode of step-by-step residual decomposition completes the learning process of gradually progressing from large-scale receptive fields to small-scale receptive fields, thereby fully exploring the inherent laws of the input data and improving the accuracy of the prediction results.

[0067] In the preprocessed data, the historical charging station utilization data of the 96 time points before the prediction period (the time point interval is 15 minutes, a total of 24 hours) and the corresponding temperature data, visibility data and date data are taken. There are 4 types of features in total, that is, the input of the hierarchical decomposition model is a sequence with a length of 96 and a dimension of 4.

[0068] First, the input sequence of the model is sent to the deep information fusion layer to obtain the information fusion sequence. The network structure of the deep information fusion layer is as follows: Figure 3 As shown in the figure, the deep information fusion layer inserts a randomly initialized marker sequence based on the standard normal distribution at the first position of the input sequence. Its function is: when using the attention encoding method for information fusion, since the marker sequence does not contain any biased information and is independent of the original input data, the sequence at the same position as the marker sequence is extracted from the attention encoding output sequence for subsequent processing, which can ensure that the information fusion sequence output by the deep information fusion layer is global and reliable.

[0069] Among them, the attention encoding method in the deep information fusion layer is as follows: first, the input sequence containing the logo sequence with a length of 96 and a dimension of 5 is multiplied by three different parameter matrices of size 5*5 to obtain the query matrix Q, the key matrix K and the value matrix V respectively, then the query matrix Q and the key matrix K are multiplied to obtain the attention matrix, and then the attention matrix is ​​normalized by the softmax function, and then the normalized attention matrix is ​​multiplied by the value matrix V to obtain the output sequence of the attention encoding method. The attention encoding process is expressed by the formula:

[0070] O = Softmax(Q·K T )·V

[0071] Among them, O represents the output sequence of the attention encoding method, Q, K, V represent the query matrix, key matrix and value matrix respectively.

[0072] The attention encoding method can learn the correlation between different sequences through sufficient data interaction between input sequences, and the introduction of the marker sequence enables it to obtain the universal correlation relationship between the original input sequences, thus completing the information fusion of the input data.

[0073] Finally, the subsequence at the first position of the sequence is extracted and sent to the fully connected layer to further extract deep information, thus obtaining an information fusion sequence with a length of 96 output by the deep information fusion layer.

[0074] Then the information fusion sequence is sent to the residual decomposition prediction layer to obtain the final prediction result. The time length of the prediction result is set to 24 (the time interval is 15 minutes, a total of 6 hours).

[0075] The network structure diagram of the residual decomposition prediction layer is as follows Figure 4 As shown. In this embodiment, the residual decomposition prediction layer is composed of 3 hierarchical components, and the output of each hierarchical component includes two parts: a hierarchical prediction sequence (length 24) and an inter-layer residual sequence (length 96). The input of each hierarchical component is the hierarchical residual sequence output by the previous hierarchical component (the input of hierarchical component 1 is the information fusion sequence). Finally, the hierarchical prediction sequences output by all hierarchical components and the inter-layer residual sequence reconstructed by the fully connected layer output by the last layer are added together to obtain the final prediction result.

[0076] In this embodiment, the hierarchical component is composed of 4 decomposition modules, and the output of each decomposition module includes two parts: a module prediction sequence (length 24) and a module backtracking sequence (length 96). The input of each decomposition module is the residual of the input sequence of the previous decomposition module and the module backtracking sequence output by the previous decomposition module (the input of decomposition module 1 is the corresponding hierarchical input sequence). Finally, the module prediction sequences output by all decomposition modules are added together to obtain the hierarchical prediction sequence output by the hierarchical component; and the residual of the input of the last decomposition module and the module backtracking sequence output by the decomposition module is the inter-layer residual sequence output by the hierarchical component.

[0077] The decomposition module is the basic component of the residual decomposition prediction layer, which is mainly composed of one-dimensional convolution, multi-layer perceptron and interpolation function. The input of the decomposition module is first downsampled by one-dimensional convolution to achieve the perception of the specific receptive field range of the module input sequence; then the downsampled sequence is passed through two different multi-layer perceptrons respectively, so as to reconstruct it into two sets of parameter sequences θ for subsequent interpolation operations. f and θ b Finally, the above two sets of parameter sequences are brought into the interpolation function, and the two sets of parameter sequences are respectively extended to the prediction length 24 of the model output and the length 96 of the model input, that is, the module prediction sequence and the module backtracking sequence are obtained.

[0078] Among them, the downsampling operation sets different sampling multiples according to the different positions of the hierarchical components, specifically: the closer the hierarchical component is to the front, the greater the downsampling multiple of the decomposition module therein. The implementation method of downsampling is: downsampling is achieved by controlling the step size of the convolution kernel in the one-dimensional convolution. The larger the step size, the greater the downsampling multiple. In this embodiment, the downsampling multiples in the decomposition modules of the hierarchical components 1, 2, and 3 are 16, 8, and 4, respectively. In order to reduce the amount of calculation of the model, the interpolation function of this embodiment selects the simplest nearest neighbor interpolation method. The interpolation multiples of the module prediction sequence and the module backtracking sequence in the decomposition modules of the hierarchical components 1, 2, and 3 are 4, 2, 1 and 16, 8, and 4, respectively.

[0079] The architecture of the residual decomposition prediction layer mainly realizes two functions. First, by controlling the different downsampling multiples in the decomposition module, the receptive field in the module input sequence is gradually refined, that is, the granularity of the prediction sequence is continuously improved; second, by repeatedly performing residual re-decomposition between decomposition modules, progressive learning composed of multiple hierarchical components is achieved, that is, the subsequent modules learn what the previous modules have not learned. Based on the above two points, the residual decomposition prediction layer realizes the improvement of the overall prediction performance of the level-by-level decomposition model.

[0080] After completing the construction of the level-by-level decomposition prediction model, the mean square error is used as the loss function of the model, and its formula is as follows:

[0081]

[0082] Among them, Loss is the loss function value, y t is the predicted value of the charging station utilization rate at time t, is the actual value of the charging station utilization rate at time t. The value range of t is 1 to 24, representing the next 24 time points (with a time interval of 15 minutes and a total of 3 hours).

[0083] Through the gradient descent algorithm, the parameters in the model will be continuously iteratively updated, so that the loss function value will continue to decrease. When the loss function value hardly changes, it means that the prediction model has converged, the training process is over, and the model parameters are fixed.

[0084] Step 4: Use the model to predict future charging station utilization:

[0085] The historical charging station utilization data, weather data and date data of the 96 time points before the prediction period are input into the step-by-step decomposition model, and a prediction sequence of length 24 is output, i.e., the predicted value of charging station utilization in the next 3 hours (forecasted every 15 minutes).

[0086] In other embodiments, the present invention also provides a charging station utilization prediction system based on a hierarchical decomposition model, comprising a memory and a processor; wherein the memory stores a computer program, and when the program is executed by the processor, the above-mentioned charging station utilization prediction method based on the hierarchical decomposition model can be implemented.

[0087] In other embodiments, the present invention further provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the above-mentioned charging station utilization prediction method based on the step-by-step decomposition model.

[0088] The parts not involved in the present invention are the same as the prior art or are implemented by using the prior art.

[0089] The above contents are further detailed descriptions of the present invention in combination with specific implementation methods, and it cannot be determined that the specific implementation of the present invention is limited to these descriptions. For ordinary technicians in the technical field to which the present invention belongs, several simple deductions or substitutions can be made without departing from the concept of the present invention, which should be regarded as falling within the protection scope of the present invention.

Claims

1. A charging station utilization prediction method based on a step-by-step decomposition model, characterized in that: include Step S100: Collect historical charging station utilization data, as well as corresponding weather data and date data as raw data; Step S200: pre-processing the original data; Step S300: constructing a step-by-step decomposition prediction model for charging station utilization prediction and training the step-by-step decomposition prediction model based on original data, wherein the step-by-step decomposition prediction model includes: a deep information fusion layer and a residual decomposition prediction layer; The original data is input into the deep information fusion layer as input data. The deep information fusion layer uses the attention encoding method to deeply extract and fuse the coupling information in the input data, obtain the information fusion sequence and input it into the residual decomposition prediction layer; the working method of the deep information fusion layer is: take the historical charging station utilization data of T time points before the prediction period, as well as the corresponding weather data and date data, and assume that there are N types of features in total, that is, the input sequence of the step-by-step decomposition prediction model is a sequence with a length of T and a dimension of N; send the input sequence into the deep information fusion layer to obtain the information fusion sequence; specifically including: First, the deep information fusion layer inserts a randomly initialized marker sequence based on the standard normal distribution at the first position of the input sequence; Secondly, the deep information fusion layer uses the attention encoding method for information fusion. Specifically, the input sequence containing the logo sequence with a length of T and a dimension of N+1 is multiplied by three different parameter matrices of size (N+1)*(N+1) to obtain the query matrix Q, the key matrix K and the value matrix V respectively. Then, the query matrix Q and the key matrix K are multiplied to obtain the attention matrix. The attention matrix is ​​then normalized by the softmax function. The normalized attention matrix is ​​then multiplied by the value matrix V to obtain the output sequence of the attention encoding method. Finally, the deep information fusion layer extracts the subsequence at the first position of the attention encoding method output sequence and sends it to the fully connected layer to further extract deep information, obtaining an information fusion sequence of length T output by the deep information fusion layer; The residual decomposition prediction layer uses the residual decomposition method to decompose the input data of the current layer step by step, and obtains prediction sequences and residual sequences of different fine-grained levels at each level. Finally, the prediction sequences of all levels and the residual sequence of the last layer through the fully connected layer are added to obtain the final prediction result; the working method of the residual decomposition prediction layer is: The residual decomposition prediction layer consists of multiple hierarchical components. The output of each hierarchical component includes two parts: the hierarchical prediction sequence and the inter-layer residual sequence. The length of the hierarchical prediction sequence is H, and the length of the inter-layer residual sequence is T. The input of each layer component is the layer residual sequence output by the previous layer component. Finally, the layer prediction sequence output by all layer components and the inter-layer residual sequence reconstructed by the fully connected layer output by the last layer are added together to obtain the final prediction result. The hierarchical component consists of multiple decomposition modules. The output of each decomposition module includes two parts: module prediction sequence and module backtracking sequence. The length of the module prediction sequence is H, and the length of the module backtracking sequence is T. The input of each decomposition module is the residual of the input sequence of the previous decomposition module and the module traceback sequence output by the previous decomposition module. Finally, the module prediction sequences output by all decomposition modules are added together to obtain the hierarchical prediction sequence output by the hierarchical component. The residual of the input of the last decomposition module and the module traceback sequence output by the decomposition module is the inter-layer residual sequence output by the hierarchical component. The decomposition module is the basic unit of the residual decomposition prediction layer, which consists of one-dimensional convolution, multi-layer perceptron and interpolation function. The input of the decomposition module is first downsampled by one-dimensional convolution, and then the downsampled sequence is passed through two different multi-layer perceptrons to reconstruct it into two sets of parameter sequences θ for subsequent interpolation operations. f and θ b ; Finally, the above two sets of parameter sequences are brought into the interpolation function, and the two sets of parameter sequences are respectively extended to the prediction length H of the model output and the length T of the model input, that is, the module prediction sequence and the module backtracking sequence are obtained; Step S400: inputting historical charging station utilization data, weather data and date data before the time period to be tested into the trained step-by-step decomposition prediction model to obtain a predicted value of charging station utilization within a preset time in the future.

2. The method for predicting charging station utilization rate based on a level-by-level decomposition model according to claim 1, characterized in that: In step S100, the weather data selects two factors that have a greater impact on travel, namely, temperature data and visibility data; Date data is accurate to the month and day, and holiday labels are added after them; the month and day are encoded with 12 and 31 digits respectively, and the holiday labels are encoded with the corresponding number of digits according to the number of holiday types; Assuming there are 10 holidays in total, the number of digits for the unique hot encoding of holidays is 10; Concatenate the unique hot codes of month, date, and holiday in sequence and add 0 after them so that the length of the date data is consistent with the length of the time series subsequently fed into the model.

3. The method for predicting charging station utilization rate based on a level-by-level decomposition model according to claim 1, characterized in that: In the step S200, data preprocessing includes data cleaning and normalization operations; Data cleaning includes correcting outliers in utilization data and weather data in the original data and filling in missing values; After completing the outlier correction and missing value filling of utilization data and weather data, normalization operations are performed on each type of data.

4. The method for predicting charging station utilization rate based on a level-by-level decomposition model according to claim 1, characterized in that: In the decomposition module, the downsampling operation sets different sampling multiples according to the different hierarchical component positions. Specifically, the more the hierarchical component is located in the front, the greater the downsampling multiple of the decomposition module in it. The implementation method of downsampling is: downsampling is achieved by controlling the step size of the convolution kernel in the one-dimensional convolution. The larger the step size, the greater the downsampling multiple.

5. The method for predicting charging station utilization rate based on a level-by-level decomposition model according to claim 1, characterized in that: In the decomposition module, the interpolation function uses nearest neighbor interpolation, linear interpolation or polynomial interpolation.

6. The method for predicting charging station utilization rate based on a level-by-level decomposition model according to claim 1, characterized in that: In step S300, the mean square error is used as the loss function of the step-by-step decomposition prediction model. Through the gradient descent algorithm, the parameters in the model will be continuously iteratively updated, so that the loss function value will continue to decrease. When the loss function value hardly changes, it means that the prediction model has converged, the training process is completed, and the parameters of the model are fixed.

7. A charging station utilization prediction system based on a step-by-step decomposition model, characterized in that: including memory and processor; Wherein, the memory stores a computer program, and when the program is executed by the processor, it can implement the charging station utilization prediction method based on the level-by-level decomposition model as described in any one of claims 1 to 6.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method for predicting the utilization rate of a charging station based on a level-by-level decomposition model described in any one of claims 1 to 6 is implemented.

Citation Information

Patent Citations

  • Patrol inspection method, device and system for charging stations

    CN106845647A

  • Multi-frame image deblurring system based on ER network and method thereof

    CN112200752A

  • Power load prediction method based on decomposition network

    CN115600640A

  • Method for realizing a multi-channel convolutional recurrent neural network EEG emotion recognition model using transfer learning

    US20230039900A1