An electric vehicle load prediction method and system based on similar day division
The electric vehicle load forecasting method based on similar day division and correlation analysis solves the problem of high difficulty in electric vehicle load forecasting, and realizes stable dispatch and safe operation of the power grid.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHINA ELECTRIC POWER RESEARCH INSTITUTE CO LTD
- Filing Date
- 2026-04-21
- Publication Date
- 2026-07-14
AI Technical Summary
Electric vehicle loads are complex and variable, and the uneven distribution of charging facilities makes it difficult to predict electric vehicle loads, which affects the stable operation of the power system.
The electric vehicle load forecasting method based on similar day division uses the K-means algorithm and XGBoost model, combined with meteorological characteristics and grid pricing information, to divide electric vehicle load into similar days and perform correlation analysis, and construct a forecasting model.
This improves the accuracy of electric vehicle load forecasting, provides a guarantee for stable grid dispatch, and ensures the safe and stable operation of the power system.
Smart Images
Figure CN122393916A_ABST
Abstract
Description
Technical Field
[0001] This invention generally relates to the field of big data analytics. More specifically, this invention relates to a method and system for predicting electric vehicle load based on similar day division. Background Technology
[0002] Driven by the dual-carbon goals and energy transition trend, new energy vehicles have experienced rapid development. According to reports from relevant institutions, the number of pure electric vehicles in my country has surged from 3.81 million in 2019 to 20.41 million in 2024, representing a growth rate of 435.70%. However, issues such as insufficient and irrational distribution of charging infrastructure make electric vehicle loads complex and variable, far more difficult to predict than residential loads. Therefore, accurate prediction of electric vehicle loads is particularly necessary to provide a basis for power dispatch and ensure the stable operation of the power system. Summary of the Invention
[0003] To at least address the technical problems described in the background section, this invention proposes an electric vehicle load forecasting method and system based on similar day division. This invention deeply integrates the effects of meteorological fluctuations, seasonal variations, and time-of-use pricing of the power grid. By performing correlation analysis on the characteristics of similar days and then combining the matching degree between electric vehicle load and these characteristics, it ensures stable power grid dispatch and provides a guarantee for the safe and stable operation of the power grid. Therefore, this invention provides solutions in the following aspects.
[0004] The first aspect of this invention provides a method for predicting electric vehicle load based on similar day segmentation, comprising: Step 1, collecting raw data on daily electric vehicle load and meteorological characteristics; Step 2, performing data processing and dataset segmentation based on the collected raw data on daily electric vehicle load and meteorological characteristics; Step 3, constructing a similar day segmentation model based on the daily electric vehicle load dataset; the similar day segmentation model outputs an elbow plot and calculates the silhouette coefficient using the K-means algorithm, determines the optimal number of clusters based on the obvious inflection point of the elbow plot and the maximum value of the silhouette coefficient, and then the K-means core clustering module performs clustering on the standardized daily load based on the optimal number of clusters to achieve similar day segmentation. K Step 4: Construct a correlation analysis model between the daily load of electric vehicles and the meteorological characteristics affecting the daily load; Step 5: Based on the similar day division and correlation analysis results, predict the daily load of electric vehicles; Step 6: Based on the predicted daily load of electric vehicles, construct evaluation indicators to assess the prediction effect.
[0005] In one embodiment, step 1 includes: Step 11: Collect raw data on the daily electric vehicle load of the area to be predicted to form a raw dataset of daily electric vehicle load. ,inT This indicates the total number of days for which data was collected; Step 12: For the aforementioned electric vehicle daily load dataset Record the corresponding features: seasonal dataset, monthly dataset, work / holiday dataset, daytime weather dataset, and nighttime weather dataset.
[0006] In one embodiment, step 2 includes: Step 21: By setting the size of the sliding window, process the raw dataset of daily load of electric vehicles. Outlier detection and removal are performed to obtain a dataset with outliers removed. Furthermore, for missing values, the missing data is supplemented using imputation methods based on periodic patterns to obtain the dataset. ; Step 22: Process the dataset The daily load dataset of electric vehicles in the dataset is subjected to maximum-minimum normalization to obtain a normalized daily load dataset of electric vehicles. ; Step 23: Convert the normalized daily load dataset of electric vehicles The dataset is randomly divided into training sets according to a preset ratio. and test set ,in and This indicates the number of days in the training and testing sets; Step 24: Perform one-hot encoding on the seasonal dataset, monthly dataset, work / holiday dataset, daytime weather dataset, and nighttime weather dataset from Step 12 to obtain a 32-dimensional feature vector corresponding to the daily load of electric vehicles. Then, obtain the feature sets corresponding to the daily load of the training set and the test set: Feature Training Set and feature test set ,in and Indicates the training set number Tianhe Test Collection No. Heavenly The feature values corresponding to the dimension.
[0007] In one embodiment, step 3 includes: For the dataset A K-means similarity day partitioning model is constructed, using 96 points per day as the smallest clustering unit. An optimal clustering number adaptive module is used, and after standardization by the maximum and minimum values, the model is selected from the candidate clusters. K Within the value range, the elbow diagram is output using the K-means algorithm and according to the formula. Calculate the profile coefficient, where , Indicates the average distance within the cluster. The minimum average distance between clusters is used to determine the optimal number of clusters based on the elbow plot's obvious inflection point and the maximum silhouette coefficient. Then, the K-means core clustering module performs clustering on the standardized daily loads based on the optimal number of clusters, thus dividing similar days into clusters. K kind.
[0008] In one embodiment, step 4 includes: Step 41: Construct a correlation analysis model for XGBoost, which is a serial ensemble regression model based on gradient boosting trees; Step 42: Based on the correlation analysis model, extract the XGBoost weights and count the number of times a feature is used as a split node in all decision trees. The more splits, the higher the importance score. Importance score of dimensional features ; Step 43: Assess the importance score. mark After normalization, we get Importance score of dimensional features .
[0009] In one embodiment, step 5 includes: Step 51: Regarding the division in step 32 Similar days, using formula Calculate the first The first cluster center dimensional eigenvalues, where , Indicates the first Class in Cluster center values on dimensional features Indicates the first The total number of samples contained in the class. Indicates the first The first in the class The sample at the th Values on the dimensional features, ; Step 52: According to the formula Compute dataset The Hamming distance between each daily load and each cluster center is used to determine the category of the daily load. The cluster center with the smallest distance is then identified as the category of the daily load. The daily load is categorized for each day; when predicting the daily load of electric vehicles on a certain day in the test set, the 5 closest days of the same category for the day to be predicted are obtained, with a distance of [missing value]. Regarding distance Take the reciprocal of each Normalization to units and processing are performed to obtain If the daily load at a certain moment is predicted, according to the formula... Calculate the predicted value for the day to be predicted, where , , , , This represents the normalized daily load dataset for the five most recent days; Step 53: Repeat step 52 to predict the daily load of all electric vehicles in the test set, obtaining a dataset of predicted values with clustering and importance scores added. .
[0010] In one embodiment, step 6 includes: Step 61: Based on the electric vehicle daily load forecast dataset obtained in Step 53 Combined with the electric vehicle daily load test dataset According to the mean absolute error Mean square error Root mean square error Mean absolute percentage error Construct evaluation indicators to assess the prediction results; Step 62: Analyze the dataset of predicted values. Maintain the same scale as the normalization in step 22 according to the formula. Perform inverse normalization to obtain the unnormalized predicted value dataset. A line graph comparing the actual and predicted daily load values of electric vehicles was plotted on a specific day.
[0011] A second aspect of the present invention provides an electric vehicle load forecasting system based on similar day division, which applies any of the above-described electric vehicle load forecasting methods based on similar day division.
[0012] This invention deeply integrates the effects of meteorological fluctuations, seasonal changes, and time-of-use pricing by the State Grid Corporation of China. By performing correlation analysis on similar day characteristics and then combining the matching degree between electric vehicle load and characteristics, it can ensure the stable dispatch of the power grid and provide a guarantee for the safe and stable operation of the power grid. Attached Figure Description
[0013] The above and other objects, features, and advantages of exemplary embodiments of the present invention will become readily apparent upon reading the following detailed description with reference to the accompanying drawings. In the drawings, several embodiments of the invention are illustrated by way of example and not limitation, and like or corresponding reference numerals denote like or corresponding parts, wherein: Figure 1 This illustrates the electric vehicle load prediction method of the present invention; Figure 2This illustrates the elbow method and contour coefficient selection of the optimal cluster number K according to the present invention; Figure 3 This illustrates typical daily load curves for each category of the present invention; Figure 4 This illustrates the importance of the 32-dimensional features of the present invention to the load; Figure 5 This is a graph showing a comparison between the actual and predicted values of the daily load of electric vehicles on a specific day according to the present invention. Detailed Implementation
[0014] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. It should be understood that the terms "first," "second," "third," and "fourth," etc., in the claims, specification, and drawings of this invention are used to distinguish different objects, rather than to describe a specific order. The terms "comprising" and "including" used in the specification and claims of this invention indicate the presence of the described features, integrals, steps, operations, elements, and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components, and / or collections thereof. It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in this specification and claims refers to any combination and all possible combinations of one or more of the associated listed items, and includes such combinations. As used in this specification and claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if [described condition or event] is detected" may be interpreted, depending on the context, as "once determined," "in response to determination," "once [described condition or event] is detected," or "in response to detection of [described condition or event]." The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings. In a first aspect, the present invention provides a method for predicting electric vehicle load based on similar day division. Figure 1This illustrates an electric vehicle load prediction method according to an embodiment of the present invention, comprising: Step 1: Collect raw data on daily load and meteorological characteristics of electric vehicles; Step 2: Based on the collected raw data of daily electric vehicle load and meteorological characteristics, perform data processing and dataset partitioning; Step 3: Based on the daily load dataset of electric vehicles, construct a similar day segmentation model; Step 4: Construct a correlation analysis model between the daily load of electric vehicles and the meteorological characteristics affecting the daily load; Step 5: Based on the results of similar day segmentation and correlation analysis, predict the daily load of electric vehicles; Step 6: Based on the daily load forecast results of electric vehicles, construct evaluation indicators to assess the forecasting effect.
[0015] Furthermore, the collection of raw data on daily load and meteorological characteristics of electric vehicles as described in step 1 specifically includes: Step 11: Collect raw data of daily electric vehicle load in the area to be predicted, recording every 15 minutes to form a raw dataset of daily electric vehicle load. ,in T This indicates the total number of days for which data was collected; Step 12: For the electric vehicle daily load dataset from Step 11 Record the corresponding features: Seasonal dataset , The values 0, 1, 2, and 3 correspond to spring, summer, autumn, and winter, respectively; month dataset. , 0 takes values from 1 to 12, corresponding to January to December respectively; Work / Holiday Dataset , Values 0 and 1 correspond to weekdays and holidays, respectively; daytime weather dataset. , Values 0-6 correspond to sunny, cloudy, partly cloudy, light rain, moderate rain, heavy rain, and torrential rain, respectively; Nighttime weather dataset , The values 0-6 correspond to sunny, cloudy, partly cloudy, light rain, moderate rain, heavy rain, and rainstorm, respectively. Furthermore, step 2 involves data processing and dataset partitioning based on the collected raw data of daily electric vehicle load and meteorological characteristics. Since the electric vehicle load is arranged in a time series, if the sliding window is set to 96 units of data per day, statistics can be calculated within this window to exclude outliers. Specifically: Step 21: For the raw dataset of electric vehicle daily load in Step 11 Set the size of the sliding window to The process involves sliding the window point-by-point along the time series, calculating statistics (mean and standard deviation) for each data point within the window, and using these as a benchmark to determine whether a data point in the current window is an outlier. If a data point in a window deviates from the window mean by more than three times the standard deviation (corresponding to severe anomalies), it is identified as an outlier and removed. For genuine outlier events (such as sudden changes in daily load caused by typhoons or meteorological variability caused by extreme weather), these are retained but marked as special points to avoid accidentally deleting critical information, resulting in a dataset with outliers removed. For outliers removed and missing values caused by sensor malfunctions, two methods are used to fill in the missing data based on periodic patterns: For consecutive missing values at multiple points, the median of the same period over multiple days (10 days, 5 days before and after) is used for filling; for very few missing values (one or two), nearest-neighbor filling is used—the dataset filled with the median of the previous day is used to fill in the missing values. ; Step 22: According to the formula For dataset The daily load dataset of electric vehicles in the dataset is subjected to maximum-minimum normalization, where... This represents the maximum value in the dataset to be normalized. This represents the minimum value in the dataset to be normalized. , Representing the data before and after normalization, respectively, we obtain the normalized daily load dataset for electric vehicles. ; Step 23: For the dataset from Step 22 according to The proportion is randomly divided into training sets and test set ,in and This indicates the number of days in the training and testing sets; Step 24: Perform one-hot encoding on the seasonal dataset, monthly dataset, weekday / weekday dataset, daytime weather dataset, and nighttime weather dataset from Step 12. Further, 1000, 0100, 0010, and 0001 correspond to spring, summer, autumn, and winter, respectively; 100000000000, ..., 00000000001 correspond to January to December, respectively; 10 and 01 correspond to weekday / weekday, respectively; and 1000000, ..., 0000001 correspond to sunny, cloudy, partly cloudy, light rain, moderate rain, heavy rain, and torrential rain, respectively. Then, arrange these sequences to obtain the 32-dimensional feature vector corresponding to the daily load of electric vehicles. This yields the feature sets corresponding to the daily load of the training and test sets: Feature Training Set. and feature test set ,in and Indicates the training set number Tianhe Test Collection No. Heavenly The feature values corresponding to the dimension; Furthermore, step 3, which involves constructing a similar day segmentation model based on the electric vehicle daily load dataset, specifically involves: Step 31: For the dataset A K-means similarity day partitioning model is constructed, using 96 points per day as the smallest clustering unit. An optimal clustering number adaptive module is used, and after standardization by the maximum and minimum values, the model is selected from the candidate clusters. K Within the value range, the elbow diagram is output using the K-means algorithm and according to the formula. Calculate the profile coefficient, where , Indicates the average distance within the cluster. The minimum average distance between clusters is used to determine the optimal number of clusters based on the elbow plot's obvious inflection point and the maximum silhouette coefficient. Then, the K-means core clustering module performs clustering on the standardized daily loads based on the optimal number of clusters, thus dividing similar days into clusters. K kind; Furthermore, step 4, which involves constructing a correlation analysis model between the daily load of electric vehicles and the meteorological characteristics affecting the daily load, specifically involves: Step 41: Construct the XGBoost correlation analysis model. This model is a serial ensemble regression model based on gradient boosting trees. Its core structure includes: First, using the average daily load of charging piles as the initial constant model, outputting initial predicted values; second, calculating the initial model prediction residuals, updating sample weights based on the residuals (enhancing the training priority of samples with larger residuals), and training the first decision tree to fit these residuals; then iteratively executing the residual calculation, sample weight update, and decision tree training steps, with each new decision tree fitting the cumulative residuals of all previous models; finally, after all decision trees are trained, the prediction results of each decision tree are weighted and summed using the learning rate (learning_rate), and then superimposed with the initial model prediction values to obtain the final predicted values. The model objective function is set to squared error loss to adapt to continuous daily load value prediction. Grid search combined with cross-validation is used to optimize key parameters, including the number of base learners (n_estimators), the maximum depth of a single decision tree (max_depth), the learning rate (learning_rate), the training sample sampling ratio (subsample), and the feature sampling ratio (colsample_bytree). L 1. Regularization coefficient (reg_alpha) L 2. Regularization coefficient (reg_lambda); To simplify the above steps, namely, constructing an XGBoost correlation analysis model, includes the following steps: Step S1: For each decision tree, during the construction process, each time a feature splitting node is selected, the reduction in MSE of the sample after the feature split (i.e., the improvement in prediction performance brought about by the split) is calculated.
[0016] Step S2: For a feature, sum the MSE reduction of all split nodes in all decision trees.
[0017] Step S3: Normalize the accumulated values of all features to obtain the final importance score.
[0018] The core logic is: the more a feature can reduce prediction error, the more important it is.
[0019] Step 42: Based on the correlation analysis model built in Step 41, extract the XGBoost weights. Count the number of times a feature is used as a splitting node across all decision trees; the more splits, the higher the importance score. Importance score of dimensional features ; Step 43: Importance score for step 42 mark Normalization is performed to make the sum equal to 1, resulting in... Importance score of dimensional features .
[0020] Furthermore, step 5, which involves predicting the daily load of electric vehicles based on the similarity day division and correlation analysis results, specifically includes: Step 51: Regarding the division in step 32 Similar days, respectively using formulas Calculate the first The first cluster center dimensional eigenvalues, where , Indicates the first Class in Cluster center values on dimensional features Indicates the first The total number of samples contained in the class. Indicates the first The first in the class The sample at the th Values on the dimensional features, ; Step 52: According to the formula Compute dataset The Hamming distance between each daily load and each cluster center is used to determine the category of the daily load. The cluster center with the smallest distance is then identified as the category of the daily load. The daily load is categorized for each day. When predicting the daily electric vehicle load for a given day in the test set, the 5 closest days within the same category for that day are used to obtain the distance. Regarding distance Take the reciprocal of each Normalization to units and processing are performed to obtain If you want to predict the daily load at a certain moment, according to the formula... Calculate the predicted value for the day to be predicted, where , , , , This represents the normalized daily load dataset for the five most recent days. Step 53: Repeat step 52 to predict the daily load of all electric vehicles in the test set, obtaining a dataset of predicted values with clustering and importance scores added. ; Furthermore, step 6 involves constructing evaluation indicators based on the daily load forecast results of electric vehicles to assess the forecasting effect, specifically as follows: Step 61: Based on the electric vehicle daily load forecast dataset obtained in Step 53 Combined with the electric vehicle daily load test dataset According to the mean absolute error Mean square error Root mean square error Mean absolute percentage error Construct evaluation indicators to assess the prediction results; Step 62: Analyze the dataset of predicted values. Maintain the same scale as the normalization in step 22 according to the formula. Perform inverse normalization to obtain the unnormalized predicted value dataset. A line graph comparing the actual and predicted daily load values of electric vehicles was plotted on a specific day.
[0021] In a specific embodiment of the present invention, the electric vehicle load prediction method based on similar day division is specifically implemented as follows: Step 1: Collect raw data on daily electric vehicle load in a typical region of China from 2023 to 2025. Record a set every 15 minutes, recording the seasonal dataset corresponding to the geographical location on the corresponding date. Monthly dataset Is it a weekday dataset? Daytime weather dataset Nighttime weather dataset .
[0022] Step 2: Set the amount of data for one day as the size of the sliding window, perform outlier detection and removal on the original data, and obtain the dataset with outliers removed. For missing values, the dataset is obtained by using imputation based on periodic patterns to fill in the missing data. A total of 62 discontinuous daily load data points for electric vehicles were collected. The dataset... The daily load dataset of electric vehicles in the dataset is subjected to max-min normalization to obtain a standardized dataset. Then the dataset according to The proportion is randomly divided into training sets and test set Then, one-hot encoding is performed on the feature dataset: 1000, 0100, 0010, 0001 correspond to spring, summer, autumn, and winter, respectively; 100000000000, ..., 000000000001 correspond to January to December, respectively; 10 and 01 correspond to whether it is a working day or not, respectively; 1000000, ..., 0000001 correspond to sunny, cloudy, partly cloudy, light rain, moderate rain, heavy rain, and rainstorm, respectively, thus obtaining the training set of the feature dataset. and test set .
[0023] Step 3: Construct a K-means clustering model and divide the constructed continuous feature dataset into groups based on the similarity of the days. Inputting the data into the model, and selecting the optimal cluster number K using the elbow method and silhouette coefficients, please refer to the appendix. Figure 2 See the appendix for typical daily load curves for each category. Figure 3 The optimal number of similar day clusters is determined using inertia and silhouette coefficient. Step 4: Construct the XGBoost correlation analysis model. The hyperparameters of the XGBoost model are shown in Table 1. Extract this attribute value as the original importance score for each feature. Then, the original importance scores are normalized to obtain the importance scores for each feature. The specific scores are as follows: Figure 4 As shown.
[0024] Table 1. Hyperparameters of the XGBoost model Step 5: For the two categories of similar days identified in Step 3, calculate the cluster centers using the formula, and combine this with the importance score to calculate the dataset. Determine the category of the day's load to be predicted based on the cluster center with the smallest distance. The daily categories are defined. Taking the prediction of daily electric vehicle load on April 9, 2023 as an example, the prediction process is explained. Based on the formula for calculating Hamming distance, the category with the smallest cluster center is determined as the category for April 9, 2023. Then, for the 5 days with the closest distance within the same category at the time of prediction, the distance is calculated. Regarding distance Take the reciprocal of each Normalization to units and processing are performed to obtain According to the formula Calculate the predicted values for each time period of the day to be predicted, and use the method described above to obtain a dataset of predicted values with added clustering and importance scores. .
[0025] Step 6: For the predicted value dataset Combined with the electric vehicle daily load test dataset The MAE, MSE, RMSE, and MAPE of each method were calculated. The evaluation indicators of the daily load prediction results of electric vehicles on the test set of different methods are shown in Table 2.
[0026] Table 2 Evaluation Indicators of Electric Vehicle Daily Load Prediction Results for Test Sets Using Different Methods For each method, a line graph comparing the actual and predicted daily load values of electric vehicles was plotted for a specific day. The comparison results are as follows: Figure 5 As can be seen from the image, the electric vehicle daily load prediction method that incorporates clustering and importance scoring can effectively track the changing trend of electric vehicle daily load and has a good prediction effect.
[0027] A second aspect of the present invention provides an electric vehicle load forecasting system based on similar day division, which runs the above-described electric vehicle load forecasting method based on similar day division.
[0028] While various embodiments of the invention have been shown and described herein, it will be apparent to those skilled in the art that such embodiments are provided by way of example only. Many modifications, alterations, and alternatives will occur to those skilled in the art without departing from the spirit and intent of the invention. It should be understood that various alternatives to the embodiments of the invention described herein may be employed in the practice of the invention. The appended claims are intended to define the scope of protection of the invention and therefore cover modular compositions, equivalents, or alternatives within the scope of these claims.
Claims
1. A method for predicting electric vehicle load based on similar day division, characterized in that, include; Step 1: Collect raw data on daily load and meteorological characteristics of electric vehicles; Step 2: Based on the collected raw data of daily electric vehicle load and meteorological characteristics, perform data processing and dataset partitioning; Step 3: Based on the electric vehicle daily load dataset, construct a similar day partitioning model. This model outputs an elbow plot using the K-means algorithm and calculates the silhouette coefficient. The optimal number of clusters is determined based on the obvious inflection points in the elbow plot and the maximum value of the silhouette coefficient. Then, the K-means core clustering module performs clustering on the standardized daily load based on the optimal number of clusters, thus partitioning the similar days into similar groups. K kind; Step 4: Construct a correlation analysis model between the daily load of electric vehicles and the meteorological characteristics affecting the daily load; Step 5: Based on the results of similar day division and correlation analysis, predict the daily load of electric vehicles; Step 6: Based on the daily load forecast results of electric vehicles, construct evaluation indicators to assess the forecast effect.
2. The electric vehicle load forecasting method based on similar day division according to claim 1, characterized in that, Step 1 includes: Step 11: Collect raw data on the daily electric vehicle load of the area to be predicted to form a raw dataset of daily electric vehicle load. ,in T This indicates the total number of days for which data was collected; Step 12: For the aforementioned electric vehicle daily load dataset Record the corresponding features: seasonal dataset, monthly dataset, work / holiday dataset, daytime weather dataset, and nighttime weather dataset.
3. The electric vehicle load forecasting method based on similar day division according to claim 1, characterized in that, Step 2 includes: Step 21: By setting the size of the sliding window, process the raw dataset of daily load of electric vehicles. Outlier detection and removal are performed to obtain a dataset with outliers removed. Furthermore, for missing values, the missing data is supplemented using imputation methods based on periodic patterns to obtain the dataset. ; Step 22: Process the dataset The daily load dataset of electric vehicles in the dataset is subjected to maximum-minimum normalization to obtain a normalized daily load dataset of electric vehicles. ; Step 23: Convert the normalized daily load dataset of electric vehicles The dataset is randomly divided into training sets according to a preset ratio. and test set ,in and This indicates the number of days in the training and testing sets; Step 24: Perform one-hot encoding on the seasonal dataset, monthly dataset, work / holiday dataset, daytime weather dataset, and nighttime weather dataset from Step 12 to obtain a 32-dimensional feature vector corresponding to the daily load of electric vehicles. Then, obtain the feature sets corresponding to the daily load of the training set and the test set: Feature Training Set and feature test set ,in and Indicates the training set number Tianhe Test Collection No. Heavenly The feature values corresponding to the dimension.
4. The electric vehicle load forecasting method based on similar day division according to claim 1, characterized in that, Step 3 includes: For the dataset A K-means similarity day partitioning model is constructed, using 96 points per day as the smallest clustering unit. An optimal clustering number adaptive module is used, and after standardization by the maximum and minimum values, the model is selected from the candidate clusters. K Within the value range, the elbow diagram is output using the K-means algorithm and according to the formula. Calculate the profile coefficient, where , Indicates the average distance within the cluster. The minimum average distance between clusters is used to determine the optimal number of clusters based on the elbow plot's obvious inflection point and the maximum silhouette coefficient. Then, the K-means core clustering module performs clustering on the standardized daily loads based on the optimal number of clusters, thus dividing similar days into clusters. K kind.
5. The electric vehicle load forecasting method based on similar day division according to claim 4, characterized in that, Step 4 includes: Step 41: Construct a correlation analysis model for XGBoost, which is a serial ensemble regression model based on gradient boosting trees; Step 42: Based on the correlation analysis model, extract the XGBoost weights and count the number of times a feature is used as a split node in all decision trees. The more splits, the higher the importance score. Importance score of dimensional features ; Step 43: Assess the importance score. mark After normalization, we get Importance score of dimensional features .
6. The electric vehicle load forecasting method based on similar day division according to claim 5, characterized in that, Step 5 includes: Step 51: Regarding the division in step 32 Similar days, using formula Calculate the first The first cluster center dimensional eigenvalues, where , Indicates the first Class in Cluster center values on dimensional features Indicates the first The total number of samples contained in the class. Indicates the first The first in the class The sample at the th Values on the dimensional features, ; Step 52: According to the formula Compute dataset The Hamming distance between each daily load and each cluster center is used to determine the category of the daily load. The cluster center with the smallest distance is then identified as the category of the daily load. The daily load is categorized for each day; when predicting the daily load of electric vehicles on a certain day in the test set, the 5 closest days of the same category for the day to be predicted are obtained, with a distance of [missing value]. Regarding distance Take the reciprocal of each Normalization to units and processing are performed to obtain If the daily load at a certain moment is predicted, according to the formula... Calculate the predicted value for the day to be predicted, where , , , , This represents the normalized daily load dataset for the five most recent days; Step 53: Repeat step 52 to predict the daily load of all electric vehicles in the test set, obtaining a dataset of predicted values with clustering and importance scores added. .
7. The electric vehicle load forecasting method based on similar day division according to claim 6, characterized in that, Step 6 includes: Step 61: Based on the electric vehicle daily load forecast dataset obtained in Step 53 Combined with the electric vehicle daily load test dataset According to the mean absolute error Mean square error Root mean square error Mean absolute percentage error Construct evaluation indicators to assess the prediction results; Step 62: Analyze the dataset of predicted values. Maintain the same scale as the normalization in step 22 according to the formula. Perform inverse normalization to obtain the unnormalized predicted value dataset. A line graph comparing the actual and predicted daily load values of electric vehicles was plotted on a specific day.
8. An electric vehicle load forecasting system based on similar day division, characterized in that, The electric vehicle load forecasting method based on similar day division as described in any one of claims 1-7 is applied.