Electric Vehicle Charging Load Prediction Method, Device, Electronic Equipment and Storage Medium

Abstract

The present disclosure belongs to the technical field of load forecasting, and provides an electric vehicle charging load forecasting method, device, electronic device and storage medium. The method includes: collecting historical load data, historical meteorological data and historical date data of each electric vehicle charging station; generating time feature data and meteorological feature data according to the historical load data, historical date data and historical meteorological data; determining regional feature data based on the location of each electric vehicle charging station; obtaining each first IMF sequence component data by using variational mode decomposition based on the historical load data; combining the time feature data, regional feature data, meteorological feature data and each first IMF sequence component data to generate each second IMF component data; inputting each second IMF component data into an LSTM-Transformer model to obtain each IMF component prediction result; summing up each IMF component prediction result to obtain an electric vehicle charging load prediction result. The present disclosure improves the accuracy of charging load forecasting.

Description

Technical Field

[0001] The present disclosure belongs to the technical field of load forecasting, and particularly relates to a method, device, electronic device and storage medium for predicting electric vehicle charging load. Background Art

[0002] With the rapid growth of the global electric vehicle market, the impact of electric vehicle charging load on the power grid has become increasingly significant. The charging demand of electric vehicles not only directly increases the power grid load, but also makes the power grid load curve more complex and dynamic. Especially during peak charging hours, a large number of charging behaviors may cause excessive pressure on the power grid and even lead to power outages in local areas. Therefore, it is necessary to predict the electric vehicle charging load to ensure the stable operation of the power grid and optimize the power grid dispatching.

[0003] Traditional methods for predicting electric vehicle charging load include using time series models for prediction and regression methods based on statistics, etc. Time series models such as Autoregressive Integrated Moving Average Model (ARIMA) and Seasonal Autoregressive Integrated Moving Average (SARIMA). The above-mentioned methods for predicting electric vehicle charging load mainly rely on historical data and single-dimensional features. When dealing with complex non-linear relationships and high-dimensional, multi-modal data, they cannot effectively capture the dynamic characteristics of electric vehicle charging behavior changing with time and environment, and it is difficult to integrate various factors affecting the charging load, resulting in large prediction errors in extreme scenarios such as peak load and spike load. Therefore, how to improve the prediction accuracy of electric vehicle charging load is an urgent problem to be solved in this field. Summary of the Invention

[0004] To solve the above problems, the present disclosure provides a method, device, electronic device and storage medium for predicting electric vehicle charging load, aiming to improve the prediction accuracy of electric vehicle charging load.

[0005] To achieve the above object, the present disclosure mainly provides the following technical solutions:

[0006] In a first aspect, the present disclosure provides a method for predicting electric vehicle charging load, including:

[0007] Collect historical load data, historical meteorological data, and historical date data of each electric vehicle charging station;

[0008] Generate time feature data according to the historical load data and the historical date data;

[0009] Generate meteorological feature data according to the historical load data and the historical meteorological data;

[0010] Determine regional feature data based on the locations of each electric vehicle charging station;

[0011] Generate a charging load time series according to the historical load data;

[0012] Decompose the charging load time series through variational mode decomposition to obtain multiple original IMF sequence component data;

[0013] Process multiple pieces of the original IMF sequence component data using a dynamic sliding window to obtain each first IMF sequence component data;

[0014] Combine the time feature data, the regional feature data, the meteorological feature data, and each of the first IMF sequence component data to generate each second IMF component data;

[0015] Input each second IMF component data into an LSTM-Transformer model to obtain each IMF component prediction result;

[0016] Sum up each IMF component prediction result to obtain an electric vehicle charging load prediction result.

[0017] In a second aspect, the present disclosure provides an electric vehicle charging load prediction device, and the device includes:

[0018] An acquisition unit, configured to acquire historical load data, historical meteorological data, and historical date data of each electric vehicle charging station;

[0019] A first generation unit, configured to generate time feature data according to the historical load data and the historical date data;

[0020] A second generation unit, configured to generate meteorological feature data according to the historical load data and the historical meteorological data;

[0021] A determination unit, configured to determine regional feature data based on the locations of each electric vehicle charging station;

[0022] A third generation unit, configured to generate a charging load time series according to the historical load data;

[0023] A decomposition unit, configured to decompose the charging load time series through variational mode decomposition to obtain multiple original IMF sequence component data;

[0024] A sliding unit, configured to process multiple pieces of the original IMF sequence component data using a dynamic sliding window to obtain each first IMF sequence component data;

[0025] A combination unit is configured to combine the time feature data, the geographical feature data, the meteorological feature data, and each of the first IMF sequence component data to generate each second IMF component data;

[0026] An input unit is configured to input each second IMF component data into an LSTM-Transformer model to obtain each IMF component prediction result;

[0027] A summation unit is configured to sum each IMF component prediction result to obtain an electric vehicle charging load prediction result.

[0028] On the other hand, the present disclosure also provides a storage medium for storing a computer program, wherein the computer program, when running, controls the device where the storage medium is located to execute the method in the first aspect above.

[0029] On the other hand, the present disclosure also provides an electronic device, which includes at least one processor, at least one memory connected to the processor, and a bus; wherein, the processor and the memory communicate with each other through the bus; the processor is configured to call program instructions in the memory to execute the method in the first aspect as described above.

[0030] Compared with the prior art, the present disclosure has the following advantages:

[0031] The present disclosure not only collects the historical load data of each electric vehicle charging station, but also collects or obtains relevant feature data such as historical meteorological data and historical date data related to the charging load. After performing variational mode decomposition on the historical load data, each first IMF sequence component data is obtained. Each second IMF component data is obtained by combining each first IMF sequence component data and each relevant feature data. The second IMF component data is used as the input data of the LSTM-Transformer model to obtain the corresponding each IMF component prediction result. Since the second IMF component data involves data related to the charging load such as temperature, meteorology, and date, the present disclosure effectively captures the dynamic characteristics of the electric vehicle charging behavior changing with time and environment through high-dimensional and multi-modal data, and the present disclosure improves the accuracy of the electric vehicle charging load prediction.

[0032] Other features and advantages of the present disclosure will be described in the subsequent specification, and part of them will become obvious from the specification, or will be understood by implementing the present disclosure. The objectives and other advantages of the present disclosure can be achieved and obtained through the structures pointed out in the specification, claims, and drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] To more clearly illustrate the technical solutions in the embodiments of the present disclosure or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present disclosure. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0034] Figure 1 Shows a schematic flowchart of a method for predicting the charging load of electric vehicles according to an embodiment of the present disclosure;

[0035] Figure 2 Shows a schematic flowchart of another method for predicting the charging load of electric vehicles according to an embodiment of the present disclosure;

[0036] Figure 3 Shows a schematic flowchart of generating combined feature vector data according to an embodiment of the present disclosure;

[0037] Figure 4 Shows a schematic flowchart of yet another method for predicting the charging load of electric vehicles according to an embodiment of the present disclosure;

[0038] Figure 5 Shows an architecture diagram of an LSTM-Transformer model according to an embodiment of the present disclosure;

[0039] Figure 6 Shows a schematic structural diagram of a device for predicting the charging load of electric vehicles according to an embodiment of the present disclosure;

[0040] Figure 7 Shows a schematic structural diagram of an electronic device according to an embodiment of the present disclosure. Detailed implementation manners

[0041] To make the objectives, technical solutions, and advantages of the embodiments of the present disclosure clearer, the following will clearly and completely describe the technical solutions in the embodiments of the present disclosure with reference to the accompanying drawings in the embodiments of the present disclosure. Obviously, the described embodiments are some, but not all, of the embodiments of the present disclosure. Based on the embodiments of the present disclosure, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present disclosure.

[0042] Figure 1 Shows a schematic flowchart of a method for predicting the charging load of electric vehicles according to an embodiment of the present disclosure. As Figure 1 shown, the method for predicting the charging load of electric vehicles in the embodiments of the present disclosure includes:

[0043] 101. Collect historical load data, historical meteorological data, and historical date data of each electric vehicle charging station.

[0044] Among them, the historical load data is used to record the charging load conditions of each electric vehicle charging station at different past time points. The historical meteorological data may include historical weather conditions, temperature, humidity and other data. The weather conditions may include sunny days, rainy days, snowy days, cloudy days, etc. The temperature may include the actual temperature and the perceived temperature. The historical date data may include date types, such as weekdays, weekends, and holidays.

[0045] In this step, the data collected and obtained can be processed in the form of word vectors in natural language processing. For example, historical date data such as weekdays, weekends, and holidays are regarded as words in natural language and converted into vector form through word vectors, so as to integrate these non-numerical data into subsequent model calculations.

[0046] 102. Generate time feature data according to the historical load data and the historical date data.

[0047] Among them, the time feature data is time data related to the electric vehicle charging load.

[0048] In the present disclosure, the historical load data can be statistically analyzed on a daily or hourly basis, combined with the date type in the historical date data, to generate all time feature data.

[0049] 103. Generate meteorological feature data according to the historical load data and the historical meteorological data.

[0050] Among them, the meteorological feature data is meteorological data related to the electric vehicle charging load.

[0051] In this step, the historical load data corresponding to different historical meteorological data can be statistically analyzed to determine the different effects of each meteorological factor on the charging load, so as to obtain the meteorological feature data.

[0052] 104. Determine the regional feature data based on the locations of each electric vehicle charging station.

[0053] Among them, the regional feature data is regional data related to the electric vehicle charging load.

[0054] In this step, the areas where each electric vehicle charging station is located can be divided, each area can be coded, and the regional feature data can be determined.

[0055] 105. Generate a charging load time series according to the historical load data.

[0056] In the embodiments of the present disclosure, the historical load data after preprocessing can be arranged in chronological order with time as the abscissa and the charging load value as the ordinate to construct a charging load time series. During the construction process, a unique charging load value can correspond to each time point.

[0057] 106. Decompose the charging load time series through variational mode decomposition to obtain multiple original IMF sequence component data.

[0058] After obtaining the charging load time series, the embodiments of the present disclosure perform variational mode decomposition (VMD) on the charging load time series to obtain multiple original intrinsic mode function (IMF) sequence component data. Before variational mode decomposition, global optimization of the variational mode decomposition hyperparameters can be performed to optimize the decomposition hyperparameters, thereby improving the charging load prediction accuracy.

[0059] 107. Process the multiple original IMF sequence component data using a dynamic sliding window to obtain each first IMF sequence component data.

[0060] Among them, the first IMF sequence component data is part of the original IMF sequence component data extracted through a sliding window. This part of the data is used to represent the local dynamic data of the charging load data. The embodiments of the present disclosure use this part of the data and the corresponding relevant feature data together to obtain the input data of the LSTM-Transformer model.

[0061] 108. Combine the time feature data, geographical feature data, meteorological feature data, and each first IMF sequence component data to generate each second IMF component data.

[0062] Among them, the second IMF component data is the input data of the LSTM-Transformer model.

[0063] In this step, each first IMF sequence component data can be combined with the corresponding time feature data, geographical feature data, and meteorological feature data to obtain each second IMF component data. It is also possible to use a sliding window to extract feature data from the first IMF sequence component data, such as mean values and extreme values, and then combine them with the corresponding time feature data, geographical feature data, and meteorological feature data to obtain the second IMF component data.

[0064] 109. Input each second IMF component data into the LSTM-Transformer model to obtain each IMF component prediction result.

[0065] In this step, the second IMF component data can be divided into training set data, test set data, and validation set data. After constructing an LSTM-Transformer model using the training set data, based on this LSTM-Transformer model, the prediction results of each IMF component can be obtained through the test set data and the validation set data.

[0066] 110. Sum up the prediction results of each IMF component to obtain the prediction result of the electric vehicle charging load.

[0067] In the embodiments of the present disclosure, not only the historical load data of each electric vehicle charging station is collected, but also historical meteorological data, historical date data and other relevant feature data related to the charging load are collected or obtained. After variational mode decomposition of the historical load data, each first IMF sequence component data is obtained. The second IMF component data is obtained by combining each first IMF sequence component data and each relevant feature data. The second IMF component data is used as the input data of the LSTM-Transformer model to obtain the corresponding prediction results of each IMF component. Since the second IMF component data involves data related to charging load such as temperature, meteorology, and date, the present disclosure effectively captures the dynamic characteristics of electric vehicle charging behavior changing with time and environment through high-dimensional and multimodal data, and the embodiments of the present disclosure improve the accuracy of electric vehicle charging load prediction.

[0068] To more specifically illustrate an electric vehicle charging load prediction method proposed by the present disclosure, an embodiment of another electric vehicle charging load prediction method is proposed by the present disclosure. The specific implementation steps of the embodiments of the present disclosure are as Figure 2 shown and include:

[0069] 201. Collect the historical load data, historical meteorological data, and historical date data of each electric vehicle charging station, and determine the region where each electric vehicle charging station is located.

[0070] As Figure 3 shown, after the embodiments of the present disclosure collect or obtain all data such as historical load data, historical meteorological data, and historical date data, the data will be cleaned and preprocessed. Based on the preprocessed data, each relevant feature data related to the charging load is determined respectively. The relevant feature data includes time feature data, geographical feature data, and meteorological feature data. The time feature data may include a daily cycle feature vector, the geographical feature data may include an encoding, that is, a target area identifier, and the meteorological feature data may include a meteorological feature weight. In addition, variational mode decomposition is performed on the historical load data to obtain each first IMF sequence component data. Feature vector combination is performed based on the first IMF sequence component data, time feature data, geographical feature data, and meteorological feature data.

[0071] 202. Generate time feature data, meteorological feature data, and regional feature data based on historical load data, historical meteorological data, historical date data, and the region where it is located.

[0072] In an implementable manner, the specific steps for generating time feature data based on historical load data and historical date data are as follows:

[0073] Step 1. Divide the historical load data with a daily cycle to obtain daily historical load data.

[0074] Step 2. Statistically analyze the daily historical load data to obtain a daily cycle feature vector.

[0075] In this step, the daily historical load data can be statistically analyzed in chronological order in hours to obtain a daily cycle feature vector.

[0076] The generation method of the daily cycle feature vector is specifically as follows:

[0077] Step 2a. Use the daily historical load data to statistically analyze the average charging load, maximum charging load, minimum charging load, and number of charging times per hour, and calculate the proportion of the charging load of each hour in the total daily charging load.

[0078] Step 2b. Generate a daily charging load distribution curve based on the average charging load, maximum charging load, minimum charging load, number of charging times, and proportion.

[0079] In the embodiments of the present disclosure, the peak period, trough period, and relatively stable period of the charging load in a day can be determined based on the daily charging load distribution curve.

[0080] Step 2c. Based on the daily historical load data, calculate the change rate of the charging load between adjacent hours to obtain supplementary feature data.

[0081] In the embodiments of the present disclosure, the change rate of the charging load between adjacent hours = (current hour load - previous hour load) / previous hour load, and this change rate is used to reflect the dynamic change trend of the charging load within a day.

[0082] Step 2d. Generate a daily cycle feature vector based on the daily charging load distribution curve and the supplementary feature data.

[0083] Among them, the daily cycle feature vector is time feature data described with a daily cycle.

[0084] In this step, the calculated change rate of the charging load between adjacent hours is used as a supplementary feature, which is integrated with the features included in the daily charging load distribution curve (such as the load statistical values, proportion information, and peak, trough, and stable period information of each time period, etc.) to obtain a daily cycle feature vector. The daily cycle feature vector can reflect the static distribution of the charging load within a day (the features of the daily load distribution curve) and the dynamic change trend (the feature of the change rate of the charging load).

[0085] Step 3: Generate time feature data by using the daily cycle feature vector and historical date data.

[0086] In the embodiments of the present disclosure, the daily cycle feature vector can be associated with the date type in the corresponding historical date data. The date type can be encoded. For example, the number 1 represents a working day, 2 represents a weekend, and 3 represents a holiday. The encoded date type and the daily cycle feature vector are combined together to form time feature data including a time pattern (date type) and a load change feature (daily cycle feature vector).

[0087] In an implementable manner, the specific steps for generating meteorological feature data according to historical load data and historical meteorological data are as follows:

[0088] Step 1: Based on the historical load data and historical meteorological data, determine the correlation between each meteorological factor and the charging load through data analysis.

[0089] Among them, the meteorological factors at least include temperature, humidity, and weather conditions, and the weather conditions include sunny days, rainy days, snowy days, and cloudy days.

[0090] In this step, by using the collected historical meteorological data (such as temperature, humidity, weather conditions, etc.) and the corresponding charging load data, the correlation between each meteorological factor and the charging load is analyzed and determined through the correlation analysis method.

[0091] Step 2: Based on the historical load data, determine the influence degree of different weather conditions on the charging load.

[0092] In this step, according to different weather conditions (sunny days, rainy days, snowy days, cloudy days, etc.), the historical charging load data when they occur can be respectively counted, including the average value, maximum value, minimum value, change range, etc. of the charging load. For example, count the average charging load of the charging station on all sunny days and the difference from the average value under other weather conditions. Compare the statistical data of the charging load under different weather conditions to determine the rules and differences in the charging load affected by the change of weather conditions. By calculating indicators such as the change amplitude and change frequency of the charging load under different weather conditions, determine the influence degree of each weather condition on the charging load.

[0093] Step 3: Assign weights to each meteorological factor according to the correlation and influence degree to obtain meteorological feature data.

[0094] In this step, based on the correlation between each meteorological factor and the charging load, as well as the influence degree of different weather conditions on the charging load, weights are assigned to each meteorological factor. The meteorological factors and their corresponding weights can be combined to obtain meteorological feature data. In the embodiments of the present disclosure, screening can also be performed according to the importance of meteorological factors during the weighting process to reduce the weights of unimportant meteorological factors.

[0095] In an implementable manner, the specific steps for determining regional feature data based on the locations of each electric vehicle charging station are as follows:

[0096] Step 1: Divide regions according to the locations of each electric vehicle charging station and the surrounding environment to obtain each target region.

[0097] Step 2: Assign region codes to each target region to generate each target region identifier.

[0098] Step 3: Determine each region identifier as regional feature data.

[0099] In the embodiments of the present disclosure, regional division can be performed according to the geographical location of the charging station and the surrounding environment, such as being divided into commercial areas, residential areas, industrial areas, transportation hub areas (such as around airports, railway stations, and bus stations), public service facility areas (such as around hospitals and schools), etc. A unique code is assigned to each region to obtain each target region identifier.

[0100] The embodiments of the present disclosure use time feature data, meteorological feature data, and regional feature data to achieve high-dimensional load prediction. As Figure 4 shown, the embodiments of the present disclosure can use the historical load data, time feature data, meteorological feature data, and regional feature data of electric vehicle charging stations to construct a load prediction data set for electric vehicle charging stations, and perform variational mode decomposition on the prediction data set to obtain each IMF sequence data and its residual m Finally, a charging load prediction model is obtained through a dynamic sliding time window and the LSTM-Transformer model. It is also possible to perform variational mode decomposition on the historical load data to obtain the corresponding IMF sequence data. After extracting feature data through a dynamic sliding time window, this part of the feature data is combined with time feature data, meteorological feature data, and regional feature data and then input into the LSTM-Transformer model. The following is a detailed description of the latter implementation method.

[0101] 203. Generate a charging load time series according to the historical load data, and decompose the charging load time series through variational mode decomposition to obtain multiple original IMF sequence component data.

[0102] In this step, the historical load data are arranged in chronological order to construct a charging load time series, and the charging load time series is subjected to variational mode decomposition to obtain multiple original IMF sequence component data.

[0103] The formula for variational mode decomposition is:

[0104] (1)

[0105] (2)

[0106] In the formula, is the k th sub-mode, is the charging load time series, is the impulse function, t is the time variable, is the central frequency of the

[0107] kth sub-mode, and k is the decomposition layer number.

[0108] In the embodiments of the present disclosure, variational mode decomposition is used to decompose the charging load time series into multiple sub-modes, process the noise information therein, and extract the time variation trend and frequency information of the load, so as to accurately predict the load. α t The embodiments of the present disclosure can also use the quadratic penalty factor

[0109] and the Lagrange multiplier λ(

[0110] In the formula, k is the th sub-mode, is the charging load time series, is the impulse function, t is the time variable, α is the central frequency of the t kth sub-mode, k is the decomposition layer number, is the quadratic penalty factor, and λ(

[0111]

[0112]

[0113]

[0114] ​​​​ (6)

[0115] Wherein, , and are respectively , and the Fourier transform of λ( t ), ω is the center frequency, n is the number of iterations, α is the quadratic penalty factor.

[0116] In the embodiment of the present disclosure, variational mode decomposition effectively reduces data noise and complexity, extracts modal components of different frequencies, makes the data easier to analyze and process, and improves the processing ability of the prediction model for the complex features of the original load data.

[0117] Since the decomposition layer number K and the quadratic penalty factor α have an important influence on variational mode decomposition, appropriate K and α can be selected before performing variational mode decomposition on the charging load time series, so as to further improve the accuracy of charging load prediction. The firefly optimization algorithm (Firefly Algorithm, FA) can be introduced to globally optimize the K and α hyperparameters of variational mode decomposition.

[0118] Use the firefly optimization algorithm to globally optimize the quadratic penalty factor α and the decomposition layer number K, and the specific method is as follows:

[0119] Step 1: Initialize the population, and set parameters such as the population size q, the maximum number of iterations Tmax, the attractiveness parameter β , the attractiveness decay coefficient γ, etc. Randomly initialize q fireflies, and set the parameters of the fireflies to ( ).

[0120] Step 2: Calculate the brightness, that is, calculate the objective function value. For each firefly, use the current to perform variational mode decomposition and calculate the objective function value. The formula of the objective function value is:

[0121] (7)

[0122] Wherein, is the decomposition layer number of the i-th firefly, is the quadratic penalty factor of the i-th firefly, is the charging load time series, N is the length of the charging load time series, and are respectively the weights of the minimum reconstruction error and the independence of the modal components, represents the k -th sub-mode.

[0123] Step 3. Update the position. For each pair of fireflies i , j Adjust the position according to the brightness comparison and the attraction formula, and calculate the Euclidean distance:

[0124] (8)

[0125] In the formula, is the decomposition layer of the i-th firefly, is the second penalty factor of the i-th firefly, is the decomposition layer of the j-th firefly, is the second penalty factor of the j-th firefly.

[0126] Calculate the attraction between two firefly individuals according to the distance between the two firefly individuals:

[0127] (9)

[0128] In the formula, is the attraction parameter, and γ is the attraction decay coefficient.

[0129] Calculate the new position according to the current position of the individual and the attraction between individuals:

[0130] (10)

[0131] (11)

[0132] In the formula, is the decomposition layer of the i-th firefly, is the second penalty factor of the i-th firefly, is the decomposition layer of the j-th firefly, is the second penalty factor of the j-th firefly, is the attraction between the i-th firefly and the j-th firefly, is the step size control parameter, is a random number between [0, 1].

[0133] From the perspective of the firefly optimization algorithm, in the formula 、 are the position parameters of the firefly individual i after being updated at the t moment respectively, 、 、 、 are the position parameters of the firefly individual at the current ( t -1 moment) respectively.

[0134] Step 4: Output the optimal solution. If the number of iterations reaches the maximum value , stop the search and output the position information of the individual with the highest brightness at present. Otherwise, return to Step 2, and use the information of the brightest individual obtained from formulas (10) and (11) at the end of the search as the optimal solution of the variational mode decomposition hyperparameters.

[0135] 204. Process multiple original IMF sequence component data by using a dynamic sliding window to obtain each first IMF sequence component data.

[0136] Step 1: Initialize parameters. Initialize the window length of the k th original IMF sequence component data as , and the sliding step size as s.

[0137] Among them, can make the windows overlap to retain more information. Set the standard deviation threshold and the change rate threshold to measure the degree of sequence fluctuation.

[0138] Step 2: Calculate the window statistics. Set the starting position of the window of the k th original IMF sequence component data as , the data within the window as , and calculate the standard deviation of the data within the window:

[0139] (12)

[0140] In the formula, is the mean value of the data within the k th window, and is the window length of the k th original IMF sequence component data.

[0141] Calculate the difference between the data in adjacent windows, and obtain the change rate sequence according to the ratio of the difference between the data in each adjacent window and the data in the corresponding previous window:

[0142] (13)

[0143] (14)

[0144] In the formula, is the difference between the data in the (i + 1)th window and the data in the ith window, and is the change rate sequence of the data within the window.

[0145] In order to avoid division by zero in the embodiments of the present disclosure, a small offset processing can be performed on the data, that is , is an extremely small positive number.

[0146] Step 3: Judge and update the window size. According to the set standard deviation threshold and the change rate threshold , obtain the comprehensive threshold:

[0147] (15)

[0148] In the formula, is the weight coefficient of is the weight coefficient of

[0149] Define the comprehensive index:

[0150] (16)

[0151] In the formula, is the weight coefficient of is the weight coefficient of is the standard deviation of the data within the k-th window, is the change rate of the data within the k-th window.

[0152] When < θ , it indicates that the data fluctuation within the window is small. At this time, the window can be appropriately reduced. On the contrary, when > θ , the window can be appropriately enlarged.

[0153] The updated window formula:

[0154] (17)

[0155] In the formula, w is the window length, α < 1, β > 1, and both are window scaling coefficients.

[0156] During the whole processing process, continuously repeat the steps of calculation, judgment and update until all the original IMF sequence component data are traversed, so as to dynamically adjust the window size to adapt to the fluctuation characteristics of different time periods of the IMF sequence. That is, obtain the optimal time window based on the updated window, and determine the data in the optimal time window as the first IMF sequence component data, so as to obtain the model input data by using the first IMF sequence component data.

[0157] 205. Combine the time feature data, geographical feature data, meteorological feature data and each first IMF sequence component data to generate each second IMF component data.

[0158] In the embodiments of the present disclosure, the first IMF sequence component data can be extracted from a dynamic sliding window, and the first IMF sequence component data can be calculated to obtain characteristic data such as the mean value, maximum value, and standard deviation of the first IMF sequence component data. The characteristic data is concatenated with temperature characteristic data, date characteristic data, and regional characteristic data to obtain the second IMF component data. Assume that the window feature vector obtained after extraction from the dynamic sliding window ( is the mean value, is the maximum value, min is the minimum value, is the standard deviation, r is the change rate sequence), after temperature normalization is , after date encoding is , after regional encoding is , the combined feature vector is then , and F is the second IMF component data. The embodiments of the present disclosure retain the dynamic local characteristics of the IMF sequence, and at the same time combine actual meteorological physical quantities, time information, and regional characteristics to achieve high-dimensional charging load prediction.

[0159] The specific method for generating each second IMF component data is as follows:

[0160] Step 1: Determine the mean value, maximum value, and standard deviation of the first IMF sequence component data.

[0161] Step 2: Concatenate the mean value, maximum value, standard deviation, and corresponding time characteristic data, regional characteristic data, and meteorological characteristic data to obtain a combined feature vector.

[0162] Step 3: Determine the combined feature vector as the second IMF component data.

[0163] 206. Input each second IMF component data into the LSTM-Transformer model to obtain the prediction results of each IMF component, and sum the prediction results of each IMF component to obtain the electric vehicle charging load prediction result.

[0164] The architecture diagram of the LSTM-Transformer model in the embodiments of the present disclosure is as Figure 5As shown, first, the input data is normalized to map the data to an appropriate range, and then it enters the LSTM layer. The LSTM can effectively handle the long-term dependencies in long sequence data and capture the temporal features of the data. The data output from the LSTM in the encoder part is added to the positional encoding. Then the data enters an encoder structure composed of N identical modules, where N is a positive integer. Each module contains a multi-head attention mechanism, add-and-normalize, a feed-forward neural network, and another add-and-normalize operation. The multi-head attention mechanism allows the model to capture different information in different representation subspaces and focus on the associations between different parts of the input sequence. In the decoder part, the decoder structure is similar to the encoder. The multi-head attention mechanism of the decoder will simultaneously focus on the output of the encoder and its own intermediate results to generate appropriate outputs. In the output layer, the output of the decoder first undergoes a dimensionality transformation through a linear layer and then passes through a Softmax layer to output the prediction results of each IMF component. The prediction results of each IMF component are summed to obtain the prediction result of the electric vehicle charging load.

[0165] The specific steps of inputting each second IMF component data into the LSTM-Transformer model to obtain the prediction results of each IMF component are as follows:

[0166] Step 1: Adjust the LSTM and Transformer parameters according to the optimal time window.

[0167] The specific steps of adjusting the LSTM and Transformer parameters are as follows:

[0168] Step 1a: Send the first IMF sequence component data into the embedding layer. The embedding layer adaptively allocates the embedding dimension according to the data length to ensure that data of different lengths can be reasonably encoded.

[0169] Step 1b: Add an adaptive layer. According to the length of the optimal time window, adjust the dimension of the initial hidden state of the LSTM and the number of heads and layers of the Transformer, so that the model architecture fits the characteristics of input data of different lengths.

[0170] Construct a mapping function to characterize the relationship between the key parameters of the LSTM and Transformer and the length of the optimal time window:

[0171] (18)

[0172] (19)

[0173] In the formula, , , , , are hyperparameters, is the length of the time window, is the optimal time window length, is the hidden state dimension, are the number of heads and the number of layers of the Transformer respectively.

[0174] The above hyperparameters can be determined by using the validation set data, trying different values, and observing the model performance on the validation set.

[0175] Step 1c: Feed the key parameters of LSTM and Transformer calculated according to the mapping rules in the adaptive layer to the LSTM and Transformer layers respectively.

[0176] Step 1d: Introduce an improved attention mechanism. The scaling factor scales the attention scores according to the optimal time window length to balance the calculation effects under different length windows. Its calculation formula is:

[0177] (20)

[0178] In the formula, Q is the query vector, K is the key vector, V is the value vector, is the optimal time window length.

[0179] Step 2: Divide the second IMF component data to obtain training set data and test set data.

[0180] In the embodiments of the present disclosure, the second IMF component data can also be divided into training set data, test set data, and validation set data.

[0181] Step 3: Input the training set data into the LSTM layer for processing to obtain the time series features of each IMF component.

[0182] Step 4: Use the time series features of each IMF component as the input of the Transformer model, and model the global dependence relationship of the time series features of each IMF component through the improved attention mechanism to construct the target LSTM-Transformer model.

[0183] In the embodiments of the present disclosure, the model training can use a loss function (such as mean square error) and an Adam optimizer to optimize the parameters. The early stopping method can also be used during the training process to prevent overfitting.

[0184] Step 5: Input the test set data into the target LSTM-Transformer model to obtain the prediction results of each IMF component.

[0185] The embodiments of the present disclosure utilize the dynamic sliding window technology to automatically adjust the window size according to the characteristics of each modal component, improving the accuracy and robustness of the LSTM-Transformer model in dealing with dependencies of different time scales. The LSTM-Transformer model combines the ability of LSTM to capture short-term dynamic characteristics and the ability of Transformer to model global dependencies, effectively solving the long-term dependency problem and significantly improving the prediction accuracy. Through experimental comparison with actual electric vehicle load data, the embodiments of the present disclosure are superior to traditional models such as LSTM, LSTM-Transformer, and VMD-LSTM in evaluation metrics such as mean absolute error (MAE), mean absolute percentage error (MAPE), and root mean square error (RMSE), providing more accurate results for electric vehicle charging load prediction and contributing to the optimal scheduling of smart grids and the development of the electric vehicle charging industry.

[0186] Based on the above method, the embodiments of the present disclosure provide an electric vehicle charging load prediction device for improving the accuracy of charging load prediction. The embodiments of this device correspond to the foregoing method embodiments. For the convenience of reading, the details in the foregoing method embodiments will not be elaborated one by one in this embodiment. However, it should be clear that the device in this embodiment can correspondingly implement all the contents in the foregoing method embodiments. Specifically, as Figure 6 shown, the device includes:

[0187] An acquisition unit 301, configured to acquire historical load data, historical meteorological data, and historical date data of each electric vehicle charging station;

[0188] A first generation unit 302, configured to generate time feature data according to the historical load data and the historical date data;

[0189] A second generation unit 303, configured to generate meteorological feature data according to the historical load data and the historical meteorological data;

[0190] A determination unit 304, configured to determine regional feature data based on the locations of each electric vehicle charging station;

[0191] A third generation unit 305, configured to generate a charging load time series according to the historical load data;

[0192] A decomposition unit 306, configured to decompose the charging load time series through variational mode decomposition to obtain multiple original IMF sequence component data;

[0193] A sliding unit 307, configured to process multiple pieces of the original IMF sequence component data by using a dynamic sliding window to obtain each first IMF sequence component data;

[0194] Combination unit 308, configured to combine the time feature data, the geographical feature data, the meteorological feature data, and each of the first IMF sequence component data to generate each second IMF component data;

[0195] Input unit 309, configured to input each second IMF component data into the LSTM-Transformer model to obtain each IMF component prediction result;

[0196] Summation unit 310, configured to sum each IMF component prediction result to obtain the electric vehicle charging load prediction result.

[0197] Further, the first generation unit is specifically configured to:

[0198] Divide the historical load data by day to obtain daily historical load data;

[0199] Statistically analyze the daily historical load data to obtain a daily cycle feature vector;

[0200] Use the daily cycle feature vector and the historical date data to generate time feature data.

[0201] Further, the first generation unit is also specifically configured to:

[0202] Use the daily historical load data to statistically analyze the average charging load, the maximum charging load, the minimum charging load, and the number of charging times per hour, and calculate the proportion of the total charging load per hour to the total daily charging load;

[0203] Generate a daily charging load distribution curve according to the average charging load, the maximum charging load, the minimum charging load, the number of charging times, and the proportion;

[0204] Based on the daily historical load data, calculate the change rate of the charging load between adjacent hours to obtain supplementary feature data;

[0205] Generate a daily cycle feature vector according to the daily charging load distribution curve and the supplementary feature data.

[0206] Further, the second generation unit is specifically configured to:

[0207] Based on the historical load data and the historical meteorological data, determine the correlation between each meteorological factor and the charging load through data analysis, where the meteorological factors at least include temperature, humidity, and weather conditions, and the weather conditions at least include sunny, rainy, snowy, and cloudy;

[0208] Based on the historical load data, determine the influence degree of different weather conditions on the charging load;

[0209] According to the relevance and the degree of influence, weights of each meteorological factor are assigned to obtain meteorological characteristic data.

[0210] Further, the determining unit is specifically configured to:

[0211] Perform regional division according to the locations and surrounding environments of each electric vehicle charging station to obtain each target region;

[0212] Assign a region code to each of the target regions to generate each target region identifier;

[0213] Determine each of the region identifiers as geographical characteristic data.

[0214] Further, the decomposing unit is further specifically configured to:

[0215] The formula of variational mode decomposition is:

[0216]

[0217]

[0218] In the formula, is the k th sub-mode, is the charging load time series, is the impulse function, t is the time variable, is the central frequency of the

[0219] th sub-mode, and k is the decomposition layer number; α Using the quadratic penalty factor t and the Lagrange multiplier λ(

[0220] ), the minimization problem is transformed into an unconstrained optimization problem, and the augmented Lagrangian function formula is: In the formula, k is the th sub-mode, is the charging load time series, is the impulse function, t is the time variable, α is the central frequency of the t th sub-mode, k is the decomposition layer number,

[0221] Iteratively update using the alternating direction method of multipliers, and stop the iteration after meeting the preset accuracy to obtain K original IMF sequence component data after decomposition. The expression of the original IMF sequence component data is:

[0222]

[0223]

[0224]

[0225] Wherein, , and are respectively the Fourier transforms of , and λ( t ), ω is the center frequency, n is the number of iterations, α is the quadratic penalty factor.

[0226] Furthermore, the decomposition unit is also specifically configured to:

[0227] Use the firefly optimization algorithm to globally optimize the quadratic penalty factor α and the decomposition layer number K.

[0228] Furthermore, the decomposition unit is also specifically configured to:

[0229] Randomly initialize q fireflies, and set the parameters of the fireflies to ([[]] );

[0230] Calculate the objective function value, and the formula of the objective function value is:

[0231]

[0232] Wherein, is the decomposition layer number of the i-th firefly, is the quadratic penalty factor of the i-th firefly, is the charging load time series, N is the length of the charging load time series, and are respectively the weights of the reconstruction error minimization and the independence of the modal components, represents the k -th sub-modal;

[0233] For each pair of fireflies i , j Adjust the positions according to the brightness comparison and the attraction formula, and calculate the Euclidean distance:

[0234]

[0235] Wherein, is the decomposition layer number of the i-th firefly, is the quadratic penalty factor of the i-th firefly, is the decomposition layer number of the j-th firefly, is the secondary penalty factor of the j-th firefly;

[0236] Calculate the attractiveness between two firefly individuals based on the distance between them:

[0237]

[0238] In the formula, is the attractiveness parameter, and γ is the attractiveness decay coefficient;

[0239] Calculate the new position based on the current position of the individual and the attractiveness between individuals:

[0240]

[0241]

[0242] In the formula, is the decomposition layer number of the i-th firefly, is the secondary penalty factor of the i-th firefly, is the decomposition layer number of the j-th firefly, is the secondary penalty factor of the j-th firefly, is the attractiveness between the i-th firefly and the j-th firefly, is the step size control parameter, is a random number between [0, 1].

[0243] Furthermore, the sliding unit is also specifically used for:

[0244] Initialize the window length of the k -th original IMF sequence component data as , and the sliding step size is s;

[0245] Set the starting position of the window of the k -th original IMF sequence component data as , the data within the window is , and calculate the standard deviation of the data within the window:

[0246]

[0247] In the formula, is the mean value of the data within the k -th window, is the window length of the k -th original IMF sequence component data;

[0248] Calculate the difference between the data in adjacent windows to obtain a change rate sequence:

[0249]

[0250]

[0251] Wherein, is the difference between the data in the (i + 1)-th window and the data in the i-th window, is the sequence of the change rates of the data in the window;

[0252] Set the standard deviation threshold as and the change rate threshold , and obtain the comprehensive threshold:

[0253]

[0254] Wherein, is 's weight coefficient, is 's weight coefficient;

[0255] Define the comprehensive index:

[0256]

[0257] Wherein, is 's weight coefficient, is 's weight coefficient, is the standard deviation of the data in the k-th window, is the change rate of the data in the k-th window;

[0258] When < θ , shrink the window, and when > θ , expand the window;

[0259] The updated window formula:

[0260]

[0261] Wherein, w is the window length, α < 1, β > 1, and both are window scaling coefficients;

[0262] Based on the updated window, obtain the optimal time window, and determine the data in the optimal time window as the first IMF sequence component data.

[0263] Furthermore, the combination unit is specifically used for:

[0264] Determine the mean value, maximum value, and standard deviation of the first IMF sequence component data;

[0265] Concatenate the mean value, the maximum and minimum values, the standard deviation, and the corresponding time feature data, geographical feature data, and meteorological feature data to obtain a combined feature vector;

[0266] Determine the combined feature vector as the second IMF component data.

[0267] Furthermore, the input unit is specifically configured to:

[0268] Adjust the LSTM and Transformer parameters according to the optimal time window;

[0269] Divide the second IMF component data to obtain training set data and test set data;

[0270] Input the training set data into the LSTM layer for processing to obtain the time series features of each IMF component;

[0271] Use the time series features of each IMF component as the input of the Transformer model, and model the global dependence relationship of the time series features of each IMF component through an improved attention mechanism to construct a target LSTM-Transformer model;

[0272] Input the test set data into the target LSTM-Transformer model to obtain the prediction results of each IMF component.

[0273] Furthermore, the input unit is also specifically configured to:

[0274] Send the first IMF sequence component data into the embedding layer, and the embedding layer adaptively allocates the embedding dimension according to the data length;

[0275] Add an adaptive layer to adjust the dimension of the initial hidden state of LSTM and the number of heads and layers of the Transformer according to the optimal time window length;

[0276] Construct a mapping function to represent the relationship between the key parameters of LSTM and Transformer and the optimal time window length:

[0277]

[0278]

[0279] where , , , , are hyperparameters, is the optimal time window length, is the hidden state dimension, They are the number of heads and the number of layers of the Transformer respectively;

[0280] Feed the key parameters of LSTM and Transformer calculated according to the mapping rule in the adaptive layer to the LSTM and Transformer layers respectively;

[0281] Introduce an improved attention mechanism. The scaling factor scales the attention score according to the optimal time window length to balance the calculation effects under different length windows. Its calculation formula is:

[0282]

[0283] In the formula, Q is the query vector, K is the key vector, and V is the value vector, is the optimal time window length.

[0284] Furthermore, the embodiments of the present disclosure further provide a processor, and the processor is used to run a program. Among them, when the program runs, it executes the above-mentioned Figure 1-2 method described therein.

[0285] Furthermore, the embodiments of the present disclosure further provide a storage medium, and the storage medium is used to store a computer program. Among them, when the computer program runs, it controls the device where the storage medium is located to execute the above-mentioned Figure 1-2 method described therein.

[0286] Furthermore, the embodiments of the present disclosure provide an electronic device 4, as Figure 7 shown. The device includes at least one processor 41, at least one memory 42 connected to the processor 41, and a bus 43; wherein, the processor 41 and the memory 42 communicate with each other through the bus 43; the processor 41 is used to call the program instructions in the memory 42 to execute the above-mentioned electric vehicle charging load prediction method. The device herein can be a server, a PC, a PAD, a mobile phone, etc.

[0287] Furthermore, the present disclosure also provides a computer program product, which is suitable for executing a program for initializing the inspection method steps of the network device as described above when executed on a data processing device.

[0288] Although the present disclosure has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present disclosure.

Claims

1. A method for predicting the charging load of an electric vehicle, characterized in that, The method includes: Collecting historical load data, historical meteorological data, and historical date data of each electric vehicle charging station; Generating time feature data according to the historical load data and the historical date data; Generating meteorological feature data according to the historical load data and the historical meteorological data; Determining regional feature data based on the locations of each electric vehicle charging station; Generating a charging load time series according to the historical load data; Decomposing the charging load time series by variational mode decomposition to obtain multiple original IMF sequence component data; Processing the multiple original IMF sequence component data by using a dynamic sliding window to obtain each first IMF sequence component data; Combining the time feature data, the regional feature data, the meteorological feature data, and each first IMF sequence component data to generate each second IMF component data; Inputting each second IMF component data into an LSTM-Transformer model to obtain each IMF component prediction result; Summing up each IMF component prediction result to obtain an electric vehicle charging load prediction result; The processing the multiple original IMF sequence component data by using a dynamic sliding window to obtain each first IMF sequence component data includes: Initialize the window length of the k first raw IMF sequence component data to be and the sliding step size is s; Set the starting position of the window for the k -th original IMF sequence component data to be , and the data within the window is . Calculate the standard deviation of the data within the window: In the formula, is the mean value of the data within the k th window, is the window length of the k th original IMF sequence component data; Calculating the data difference within adjacent windows to obtain a change rate sequence: In the formula, is the difference between the data in the (i + 1)-th window and the data in the i-th window, is the sequence of the change rate of the data within the window; Set the standard deviation threshold to and the change rate threshold , to obtain the comprehensive threshold: In the formula, is 's weight coefficient, is 's weight coefficient; Defining a comprehensive index: Wherein, is 's weight coefficient, is 's weight coefficient, is the standard deviation of the data within the k-th window, is the change rate of the data within the k-th window; When < θ shrink the window, and when > θ enlarge the window; An updated window formula: In the formula, w is the window length, α < 1, β > 1, and both are window scaling coefficients; Obtaining an optimal time window based on the updated window and determining the data in the optimal time window as the first IMF sequence component data.

2. The method according to claim 1, characterized in that, Generating time feature data according to the historical load data and the historical date data includes: Dividing the historical load data by day to obtain daily historical load data; Statistically analyzing the daily historical load data to obtain a daily cycle feature vector; Generating time feature data by using the daily cycle feature vector and the historical date data.

3. The method according to claim 2, wherein Statistically analyzing the daily historical load data to obtain a daily cycle feature vector includes: Using the daily historical load data to statistically analyze the average charging load, the maximum charging load, the minimum charging load, and the number of charging times per hour, and calculating the proportion of the total charging load per hour to the total daily charging load; Generating a daily charging load distribution curve according to the average charging load, the maximum charging load, the minimum charging load, the number of charging times, and the proportion; Calculating the change rate of the charging load between adjacent hours based on the daily historical load data to obtain supplementary feature data; Generating a daily cycle feature vector according to the daily charging load distribution curve and the supplementary feature data.

4. The method according to claim 1, wherein Generating meteorological feature data according to the historical load data and the historical meteorological data includes: Based on the historical load data and the historical meteorological data, determining the correlation between each meteorological factor and the charging load through data analysis, where the meteorological factors at least include temperature, humidity, and weather conditions, and the weather conditions at least include sunny, rainy, snowy, and cloudy; Based on the historical load data, determine the influence degree of different weather conditions on the charging load; According to the correlation and the influence degree, assign weights to each meteorological factor to obtain meteorological feature data.

5. The method according to claim 1, wherein Based on the locations of each electric vehicle charging station, determine regional feature data, including: Conduct regional division according to the locations of each electric vehicle charging station and the surrounding environment to obtain each target area; Assign area codes to each of the target areas to generate each target area identifier; Determine each of the area identifiers as regional feature data.

6. The method according to claim 1, characterized in that, Decompose the charging load time series through variational mode decomposition to obtain multiple original IMF sequence component data, including: The formula for variational mode decomposition is: In the formula, is the k sub-modal, is the charging load time series, is the impulse function, t is the time variable, is the center frequency of the k-th sub-modal, and k is the decomposition level; Using the quadratic penalty factor α and the Lagrange multiplier λ( t ), the minimization problem is transformed into an unconstrained optimization problem, and the augmented Lagrangian function formula is as follows: In the formula, is the k th sub-modal, is the charging load time series, is the impulse function, t is the time variable, is the center frequency of the kth sub-modal, k is the decomposition level, α is the quadratic penalty factor, λ( t ) is the Lagrange multiplier; Use the alternating direction multiplier method to iteratively update and stop the iteration after meeting the preset accuracy to obtain K original IMF sequence component data after decomposition, and the expression of the original IMF sequence component data is: In the formula, , and are respectively , and the Fourier transform of λ( t ), ω is the center frequency, n is the number of iterations, α is the quadratic penalty factor.

7. The method according to claim 6, wherein The method further includes: Use the firefly optimization algorithm to globally optimize the quadratic penalty factor α and the decomposition level K.

8. The method according to claim 7, characterized in that, Use the firefly optimization algorithm to globally optimize the quadratic penalty factor α and the decomposition level K, including: Random initialization q Fireflies, and set the parameters of the fireflies to ( ); Calculate the objective function value, and the formula for the objective function value is: wherein, is the decomposition level of the i-th firefly, is the secondary penalty factor of the i-th firefly, is the charging load time series, N is the length of the charging load time series, and are the weights of the reconstruction error minimization and the independence of the modal components, respectively, denotes the k -th sub-modal; For each pair of fireflies i , j Adjust the positions according to the brightness comparison and the attraction formula, and calculate the Euclidean distance: Wherein, is the decomposition level of the i-th firefly, is the secondary penalty factor of the i-th firefly, is the decomposition level of the j-th firefly, is the secondary penalty factor of the j-th firefly; Calculate the attraction degree between two firefly individuals according to the distance between the two firefly individuals: wherein, is the attraction parameter, and γ is the attraction attenuation coefficient; Calculate a new position according to the current position of the individual and the attraction degree between individuals: Wherein, is the decomposition level of the i-th firefly, is the quadratic penalty factor of the i-th firefly, is the decomposition level of the j-th firefly, is the quadratic penalty factor of the j-th firefly, is the attractiveness between the i-th firefly and the j-th firefly, is the step size control parameter, is a random number between [0, 1].

9. The method according to claim 1, wherein Combine the time feature data, the regional feature data, the meteorological feature data, and each of the first IMF sequence component data to generate each second IMF component data, including: Determine the mean value, the maximum value, and the standard deviation of the first IMF sequence component data; Concatenate the mean value, the maximum value, the standard deviation, and the corresponding time feature data, regional feature data, and meteorological feature data to obtain a combined feature vector; Determine the combined feature vector as the second IMF component data.

10. The method according to claim 9, characterized in that Input each second IMF component data into the LSTM-Transformer model to obtain each IMF component prediction result, including: Adjust the LSTM and Transformer parameters according to the optimal time window; Divide the second IMF component data to obtain training set data and test set data; Input the training set data into the LSTM layer for processing to obtain each IMF component time series feature; Use each of the IMF component time series features as the input of the Transformer model, and perform global dependency modeling on each of the IMF component time series features through an improved attention mechanism to construct a target LSTM-Transformer model; Input the test set data into the target LSTM-Transformer model to obtain each IMF component prediction result.

11. The method according to claim 10, characterized in that, Adjust the LSTM and Transformer parameters according to the optimal time window, including: Send the first IMF sequence component data into the embedding layer, and the embedding layer adaptively assigns embedding dimensions according to the data length; Add an adaptive layer to adjust the initial hidden state dimension of the LSTM and the number of heads and layers of the Transformer according to the optimal time window length; Construct a mapping function to represent the relationship between the key parameters of the LSTM and the Transformer and the optimal time window length: In the formula, , , , are hyperparameters, is the optimal time window length, is the hidden state dimension, are the number of heads and the number of layers of the Transformer respectively; The key parameters of LSTM and Transformer calculated according to the mapping rules in the adaptive layer are respectively fed to the LSTM and Transformer layers; An improved attention mechanism is introduced. The scaling factor scales the attention scores according to the optimal time window length to balance the calculation effects under different length windows. The calculation formula is: Where Q is the query vector, K is the key vector, and V is the value vector, is the optimal time window length.

12. An electric vehicle charging load prediction device, characterized in that, The device includes: An acquisition unit for acquiring historical load data, historical meteorological data, and historical date data of each electric vehicle charging station; A first generation unit for generating time feature data according to the historical load data and the historical date data; A second generation unit for generating meteorological feature data according to the historical load data and the historical meteorological data; A determination unit for determining regional feature data based on the locations of each electric vehicle charging station; A third generation unit for generating a charging load time series according to the historical load data; A decomposition unit for decomposing the charging load time series through variational mode decomposition to obtain a plurality of original IMF sequence component data; A sliding unit for processing the plurality of original IMF sequence component data by using a dynamic sliding window to obtain each first IMF sequence component data; A combination unit for combining the time feature data, the regional feature data, the meteorological feature data, and each of the first IMF sequence component data to generate each second IMF component data; An input unit for inputting each second IMF component data into an LSTM-Transformer model to obtain each IMF component prediction result; A summation unit for summing each IMF component prediction result to obtain an electric vehicle charging load prediction result; The sliding unit specifically is used for: Initialize the k window length of the first raw IMF sequence component data to be s ; the sliding step size is s; Set the starting position of the window for the k th original IMF sequence component data to be , and the data within the window is , and calculate the standard deviation of the data within the window: Wherein, is the mean value of the data within the k th window, is the window length of the k th original IMF sequence component data; Calculating the data difference within adjacent windows to obtain a change rate sequence: wherein, is the difference between the data in the (i + 1)-th window and the data in the i-th window, is the sequence of the change rate of the data within the window; Set the standard deviation threshold to and the change rate threshold , and obtain the comprehensive threshold: In the formula, is 's weight coefficient, is 's weight coefficient; Defining a comprehensive index: In the formula, is 's weight coefficient, is 's weight coefficient, is the standard deviation of the data within the k-th window, is the change rate of the data within the k-th window; When < θ shrink the window, and when > θ enlarge the window; The updated window formula: In the formula, w is the window length, α < 1, β > 1, and both are window scaling coefficients; Based on the updated window, an optimal time window is obtained, and the data in the optimal time window is determined as the first IMF sequence component data.

13. The device according to claim 12, wherein The first generation unit specifically is used for: Dividing the historical load data with a daily cycle to obtain daily historical load data; Statistical analysis is performed on the daily historical load data to obtain a daily cycle feature vector; Using the daily cycle feature vector and the historical date data to generate time feature data.

14. An electronic device, characterized in that, The device includes at least one processor, at least one memory connected to the processor, and a bus; wherein, the processor and the memory complete communication with each other through the bus; the processor is used to call program instructions in the memory to execute the method according to any one of claims 1-11.

15. A computer storage medium, characterized in that, The storage medium is used to store a computer program, wherein the computer program controls the device where the storage medium is located to execute the method according to any one of claims 1-11 when running.

Images