Method and system for predicting electric vehicle charging demand based on machine learning

Abstract

The application discloses a method and system for predicting electric vehicle charging demand based on machine learning, comprising: obtaining charging order data, constructing a user behavior portrait feature set based on historical charging order data, dividing the prediction space into a regular domain and a specific domain through preliminary marking and outlier samples according to the user behavior portrait, performing point prediction on the charging demand of the future period based on the regular domain, abandoning point prediction on the specific domain, and outputting a prediction confidence interval based on the historical same period deviation information of the future period, evaluating the prediction credibility according to the point prediction result, the prediction confidence interval and the current specific sample density, and obtaining differentiated charging demand prediction information based on the prediction credibility. The method uses a hierarchical prediction strategy of regular domain point prediction and specific domain historical deviation confidence interval, improves the prediction reliability of charging demand in complex scenarios, and improves the practicality and interpretability of the prediction result for power grid dispatching decisions.

Description

Technical Field

[0001] This invention relates to the field of electric vehicle charging management, and more particularly to a method and system for predicting electric vehicle charging demand based on machine learning. Background Technology

[0002] With the rapid development of the electric vehicle industry and the continuous growth of market share, large-scale charging behavior has placed higher demands on the operational safety of urban power distribution networks and the operational efficiency of public charging stations. Electric vehicle charging demand forecasting, by mining the correlation characteristics of multi-source data such as historical charging data, user behavior patterns, and spatiotemporal environment, predicts the charging load and charging capacity demand in future periods. It is a key foundation for supporting power distribution network load scheduling, charging station resource allocation optimization, and user charging guidance. The mainstream forecasting methods can be divided into methods based on statistical time series models and methods based on machine learning. Methods such as ARIMA and grey prediction fit charging demand through the historical evolution of charging, while machine learning solutions mostly adopt a globally unified modeling approach. They have certain predictive effects in conventional scenarios, but when faced with complex and irregular charging scenarios, the prediction accuracy and practicality become unstable, making it difficult to adapt to the complex power grid, charging stations, and electric vehicle charging demands.

[0003] Existing charging demand forecasting models often output single-point prediction results, failing to design differentiated forecasting strategies based on the regularity and randomness of charging behavior. This results in low reliability of the predictions, making them unsuitable for power grid dispatching and charging station operation. Furthermore, they lack sufficient analysis of user charging behavior preferences. Therefore, how to perform domain-specific modeling of user behavior and differentiate charging demand forecasting to improve prediction accuracy and reliability, and provide more practical forecasting information, has become an urgent problem to be solved. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for predicting electric vehicle charging demand based on machine learning.

[0005] To achieve the above objectives, the present invention is implemented according to the following technical solution: The first aspect of this invention provides a machine learning-based method for predicting electric vehicle charging demand, comprising: Obtain charging order data and construct a user behavior profile feature set based on historical charging order data. The profile feature set includes temporal regularity, spatial jump, energy preference, and demand stability. User behavior profiles are probabilistic soft clustering using a Gaussian mixture model. Specific users are pre-labeled based on soft attribution entropy, and outlier samples are filtered out using an isolation forest. The prediction space is divided into a regular domain and a specific domain based on the preliminary labeling and outlier samples. Based on the conventional domain, point predictions are made for charging demand in future periods, while point predictions are abandoned for special domains. Confidence intervals are output based on historical deviation information from the same period in future periods. The prediction reliability is assessed based on the point prediction results, prediction confidence interval, and current specific sample density. Based on the prediction reliability, differentiated charging demand prediction information is obtained.

[0006] Furthermore, the method for obtaining the user behavior profile feature set includes: Historical charging order data is anonymized, user-level aggregation is performed based on the anonymization identifier, and historical charging behavior statistics for each user are extracted through a preset sliding time window. The historical charging behavior statistics include the charging period, charging duration, charging station identifier, charging amount, and charging power for each order. The charging frequency distribution for each charging period is calculated based on the start and end times, and the information entropy and standard deviation of the charging frequency distribution for each period are used as the degree of time regularity. Geographical distances between charging stations are obtained based on charging station identifiers, and the number of charging stations is counted to obtain spatial jump degree. The average single charging power and the proportion of fast charging are used as energy preferences. The coefficient of variation of weekly charging frequency is calculated to obtain demand stability. Based on time regularity, spatial jump degree, energy preference and demand stability, a wide table of user-level behavior profiles is constructed to obtain the user behavior profile feature set.

[0007] Furthermore, the method for obtaining the specific user includes: The user behavior profile feature set is Z-score standardized. Based on the Bayesian information criterion, the optimal number of clusters in the Gaussian mixture model is determined by the elbow rule. The Gaussian mixture model is configured according to the optimal number of clusters. The standardized feature set is input into the Gaussian mixture model for training until the change in the log-likelihood function of adjacent iterations is less than or equal to the preset convergence threshold. The posterior probability matrix of each user sample belonging to each cluster is output. Based on the posterior probability matrix, each user sample is assigned to the cluster with the highest posterior probability to obtain the behavioral subgroup division results of each user. The information entropy of each user sample is calculated according to the posterior probability matrix as the soft attribution entropy, and the soft attribution entropy distribution of each user sample is obtained. The upper quartile and interquartile range are calculated according to the soft attribution entropy distribution using the box plot statistical method. Based on the upper quartile and interquartile range, the anomaly threshold is determined, and users whose soft attribution entropy exceeds the anomaly threshold are pre-marked as special users.

[0008] Furthermore, the method for obtaining the conventional domain and the special domain includes: Based on the behavioral subgroup segmentation results, the standardized user behavior profile feature sets within each behavioral subgroup are input into the isolated forest for anomaly detection. Outlier scores of user samples within each subgroup are output. Outlier samples are filtered by time period according to the distribution of user outlier scores. Outlier samples are sorted in descending order according to outlier scores. Pre-labeled specific users and outlier samples are deduplicated and merged to obtain a set of specific samples. The remaining non-specific samples are used as regular samples. Based on the two types of samples, the prediction space is divided into a regular domain and a specific domain. The regular domain includes users and charging scenarios corresponding to regular samples, and the specific domain includes users and charging scenarios corresponding to specific samples.

[0009] Furthermore, the method for obtaining the point prediction result includes: Based on the conventional domain, the charging demand for future periods is regressed and predicted using the first prediction model to obtain point prediction results, which are used as the baseline prediction values ​​for the conventional domain. For the prediction of abandonment points in the special domain, the charging demand deviation is obtained by aggregating the charging power difference between historical special periods and regular periods for each special user in the special domain. The sum of the baseline prediction value of the regular domain and the charging demand deviation is used as the prediction baseline for the special domain. The formula for calculating the charging demand deviation is as follows: ; in The amount of deviation in charging demand caused by special users. For time period Internal special user set, For user samples Historical anomalous order subsets User samples A subset of historical regular orders For orders The charging capacity, For orders The charging capacity.

[0010] Furthermore, the method for obtaining the prediction confidence interval includes: Based on the specific domain, deviation sequences relative to the baseline of the conventional domain are extracted from historical demand during the same period. A tail sample subset is determined based on the deviation sequences. The upper bound of the interval is calculated based on the tail sample mean and tail sample standard deviation. The formula for calculating the upper bound of the interval is as follows: ; ; in For deviations from the sequence exceeding The tail sample subset of quantiles, As an off-sequence, For the first deviated sequence Demand deviation during a historically unique period , where is the number of tail samples. The upper bound of the interval is... This is the tail magnification factor. The summation index within the tail sample subset; The lower bound of the interval is determined based on the quantiles of the deviation sequence, and the formula for calculating the lower bound of the interval is as follows: ; in The lower bound of the interval, For the off-sequence quantiles; Based on the specific domain prediction baseline, upper bound, and lower bound, quantile regression is performed using a second prediction model to obtain the prediction confidence interval.

[0011] Furthermore, the method for obtaining the charging demand prediction information includes: The density of specific samples is determined based on the ratio of the number of specific samples to the total number of active users in the current period, and the baseline deviation is calculated based on the degree of deviation between the specific domain prediction baseline and the regular domain baseline prediction value. The specific sample density, prediction confidence interval width, and baseline deviation are weighted and fused to obtain the prediction confidence level. Based on the numerical distribution of the prediction confidence level, quantiles are performed to obtain high confidence, medium confidence, and low confidence levels. Differentiated charging demand prediction information is output according to the confidence level, point prediction results, and prediction confidence interval. The differentiated charging demand prediction information includes scheduling early warning information.

[0012] A second aspect of the present invention provides a machine learning-based electric vehicle charging demand prediction system, comprising: Data acquisition module: used to acquire charging order data and construct a user behavior profile feature set based on historical charging order data. The profile feature set includes temporal regularity, spatial jump, energy preference, and demand stability. The module for dividing the regular and special domains is used to perform probabilistic soft clustering of user behavior profiles using a Gaussian mixture model, pre-label special users based on soft attribution entropy, filter outlier samples through an isolation forest, and divide the prediction space into a regular and special domains based on the preliminary labels and outlier samples. Point prediction module: used to make point predictions of charging demand in future periods based on the conventional domain, abandon point prediction for special domains, and output prediction confidence intervals based on historical deviation information of the same period in future periods. Demand forecasting information output module: used to evaluate the forecast credibility based on the point forecasting results, forecast confidence interval and current specific sample density, and obtain differentiated charging demand forecasting information based on the forecast credibility.

[0013] Compared with the prior art, the embodiments of the present invention have at least the following advantages or beneficial effects: This invention constructs user behavior profiles and performs two-stage anomaly identification, dividing the prediction space into a regular domain and an anomaly domain. It distinguishes between regular and random charging behaviors, avoiding prediction distortion caused by the global model treating anomalies as noise smoothing. Point prediction is performed in the regular domain to ensure baseline accuracy, while the prediction confidence interval is output based on historical deviation information in the anomaly domain. The demand fluctuation boundary under extreme scenarios is quantified by asymmetric charging demand envelope to evaluate prediction reliability, provide a tolerable reference for power grid dispatch, output differentiated charging demand prediction information, and obtain optimization suggestions. Attached Figure Description

[0014] Figure 1 This is a flowchart illustrating the steps of the electric vehicle charging demand prediction method based on machine learning in an embodiment of the present invention. Detailed Implementation

[0015] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0016] Reference Figure 1 As shown, this invention provides a machine learning-based method for predicting electric vehicle charging demand, including: Obtain charging order data and construct a user behavior profile feature set based on historical charging order data. The profile feature set includes temporal regularity, spatial jump, energy preference, and demand stability. In the actual evaluation, operational data of preset charging piles from January to June were obtained, including anonymous charging order data from approximately 5,000 users at 10 charging stations in the core area. This data covered commuter drivers and cross-regional travel drivers. Commuter drivers used fixed routes and had stable charging behavior, while cross-regional travel drivers were temporary charging users with irregular behavior. User information in the charging order data was anonymized, and no user-identifiable information was available. Data was aggregated by anonymous user ID, and a 30-day sliding time window was used to extract historical charging behavior statistics for each user. This included hourly charging time slots, charging duration, station ID, charging amount (kWh), charging power (kWh), and charging gun type (including fast and slow charging) for each order. Temporal regularity, spatial abruptness, energy preference, and demand stability were calculated. For user A, the charging time slots were concentrated between 6 PM and 8 PM, with a frequency distribution information entropy of 0.2 and a standard deviation of 0.1. Spatial abruptness showed that user A frequently visited two stations. The geographical distance between the stations is approximately 1.5km. User C's energy preference is an average single charge of 30kWh, with fast charging accounting for 20%. Demand stability is stable at 5-6 times per week, with a coefficient of variation of 0.1. User C's charging time is distributed from 0:00 to 24:00, concentrated between 0:00 and 6:00, with an information entropy of 0.8 and a standard deviation of 1.2. User C frequently visits 3 stations, distributed on the east and west sides of the core area, with an average distance of 8km. The average single charge is 45kWh, with fast charging accounting for 90%. The weekly charging frequency is stable at 15-18 times, with a coefficient of variation of 0.15. User B's charging time is distributed from 0:00 to 24:00, with an information entropy of 1.6 and a standard deviation of 2.5. User B frequently visits 8 stations, distributed throughout the core area, with an average distance of 60km. The average single charge is 65kWh, with fast charging accounting for 87%. The weekly charging frequency is stable at 2-10 times, with a coefficient of variation of 0.8. A wide table of user-level behavioral profiles is constructed to obtain the user behavioral profile feature set. User behavior profiles are probabilistic soft clustering using a Gaussian mixture model. Specific users are pre-labeled based on soft attribution entropy, and outlier samples are filtered out using an isolation forest. The prediction space is divided into a regular domain and a specific domain based on the preliminary labeling and outlier samples. In the actual evaluation, features such as temporal regularity and spatial abruptness were Z-score standardized to eliminate dimensional differences. The optimal number of clusters was determined using the BIC Bayesian information criterion and the elbow rule, with the cluster number ranging from 2 to 8. The model fitting effect and information loss corresponding to different cluster numbers were compared, and the optimal number of clusters for the local charging user group was determined to be 3 clusters, roughly corresponding to three user types: commuter, operational, and individual. A Gaussian mixture model (GMM) was configured based on the optimal number of clusters, with a preset convergence threshold of 1e-6. The standardized feature set was input into the GMM model for training, with a maximum of 100 iterations. Iterate until the log-likelihood change is less than 1e-6, then output the posterior probability of each user belonging to each cluster. Users belonging to the commuter cluster have a posterior probability greater than or equal to 0.95, the operational cluster 0.03, and the individual passenger cluster 0.02. User B has a probability of 0.33 in all three clusters, indicating a uniform posterior probability distribution. Based on the posterior probability matrix, each user sample is assigned to the cluster with the highest posterior probability, obtaining the behavioral subgroup division results. The information entropy of each user sample is calculated based on the posterior probability matrix; user A has an entropy value of 0.2, and user B has an entropy value of 1.6. The upper quartile and interquartile range are calculated based on the entropy distribution of the user samples, automatically... An upper boundary threshold for anomaly detection is generated based on entropy values. User B's entropy value exceeds the threshold, and they are pre-marked as an outlier. Based on the clustering results of Gaussian Mixture Model (GMM), the standardized user behavior profile feature sets within each behavioral subgroup are input into an isolated forest for anomaly detection, using 100 decision trees and a subsampling ratio of 0.7. Considering the strong time dependence of electric vehicle charging, the entire dataset is divided into weekday off-peak hours, weekday evening peak hours, weekend hours, and public holiday hours. Outlier scores are calculated for users within each subgroup during different time periods. Outlier scores closer to 1 indicate a higher likelihood of a standard user sample within the same group. The higher the degree of deviation in charging behavior within the same period, the higher the outlier the user D's average charging amount is (100 kWh / h), which is higher than the cluster average of 45 kWh / h. The outlier score is 0.8, and it is marked as an outlier sample. A total of 154 outlier samples were detected. The pre-labeled special user B and the outlier sample D were merged and deduplicated to obtain the special sample set. There are 320 special samples, accounting for 6.4% of the total. The rest are regular samples. Based on the two types of samples, the prediction space is divided into a regular domain and a special domain. The regular domain contains 4680 user samples, mainly corresponding to commuting needs and stable ride-hailing services. The special domain contains 320 user samples, corresponding to cross-domain and irregular charging. Based on the conventional domain, point predictions are made for charging demand in future periods, while point predictions are abandoned for special domains. Confidence intervals are output based on historical deviation information from the same period in future periods. In the actual evaluation, a training set was generated using historical charging data, time periods, and weather conditions of regular domain user samples. The LightGBM time-series regression model was then trained based on this training set to fit the charging load of the charging stations. The predicted future time period was Monday evening peak from 18:00 to 19:00 on July 1st. The model output a baseline prediction value of 1180 kWh for the regular domain, representing the basic charging load formed by a stable user group, and was used as the prediction result for the regular domain. For the special domain, single-point numerical predictions were directly abandoned. Based on 320 statistical data points of specific historical charging behavior, historical special charging periods, regular charging periods, and the average single-charge amount under both scenarios were distinguished. These were then substituted into the charging demand deviation calculation, yielding a charging demand deviation of 265 kWh caused by the special user group during the predicted time period. The baseline prediction value for the regular domain was then superimposed with the deviation value to calculate the prediction for the special domain. The baseline is 1180+265=1445kWh. Specific charging demand data corresponding to the evening peak hours of Mondays from January to June (18:00-19:00) are extracted. Based on the fluctuation difference of charging demand in the scenario relative to the conventional baseline, a historical demand deviation sequence for the same period is constructed. The sequence data are all filtered according to time-segmentation rules, retaining only valid deviation samples of the same scenario and time period, and removing irrelevant samples such as those from holidays and extreme weather to obtain the historical deviation sequence. The tail sample subset is determined based on the deviation sequence, where 1-α=0.95 is used as the upper bound quantile, and the tail amplification factor is 1.2. The upper bound of the interval is calculated to be 312, and the lower bound β is taken as the 5th quantile, with a lower bound of 85. The quantile regression model is used as the second prediction model. Based on the specific domain prediction baseline, the upper bound of the interval, and the lower bound of the interval, quantile regression is performed through the second prediction model to obtain the prediction confidence interval [1530, 1757]. The prediction reliability is assessed based on the point prediction results, prediction confidence interval, and current specific sample density. Based on the prediction reliability, differentiated charging demand prediction information is obtained.

[0017] In the actual evaluation, the number of unique users in the predicted period was divided by the total number of active users (1600), resulting in a unique sample density of 20%. The average absolute error of the historical data from the last 20 similar evening peak periods was linearly normalized to the [0,1] interval. The baseline deviation for this period was 0.23, and the interval width was 1757-1530=227kWh, with a ratio of 0.16 to the baseline of 1445. The unique sample density had a weight of 0.4, the interval width a weight of 0.3, and the baseline deviation a weight of 0.3. The calculated prediction confidence was 0.8. After quantile segmentation, low confidence was >0.8, medium confidence was 0.5-0.8, and high confidence was <0.5. A confidence level of 0.82 falls within the medium confidence level. Commuter users primarily have home charging needs, charging at fixed times and only replenishing power when the remaining battery is low. The charging time and... With low tolerance for load fluctuations, operational users have flexible daily travel times and can adjust charging times independently, showing a high tolerance for time fluctuations in charging demand. Individual users prioritize charging during off-peak hours and when charging stations are idle, flexibly arranging their charging needs based solely on external conditions. User A is a commuter with a low coefficient of variation in charging frequency and a high initial charge. User C is an individual user with greater flexibility in charging time. Therefore, the core area charging stations obtain differentiated charging demand forecast information for Monday evening peak hours from 18:00 to 19:00: "There are many conservative users, with an exceptional sample accounting for 20%, a predicted charging demand value of 1180kWh, a moderate prediction confidence level, and a prediction confidence interval of [1530, 1757]kWh. It is noted that there is a risk of temporary charging load increases during the evening peak hours, and it is recommended to configure off-peak charging in advance and reserve about 10% of the power load redundancy to cope with temporary surges in exceptional charging demand."

[0018] In this embodiment, the method for obtaining the user behavior profile feature set includes: Historical charging order data is anonymized, user-level aggregation is performed based on the anonymization identifier, and historical charging behavior statistics for each user are extracted through a preset sliding time window. The historical charging behavior statistics include the charging period, charging duration, charging station identifier, charging amount, and charging power for each order. The charging frequency distribution for each charging period is calculated based on the start and end times, and the information entropy and standard deviation of the charging frequency distribution for each period are used as the degree of time regularity. Geographical distances between charging stations are obtained based on charging station identifiers, and the number of charging stations is counted to obtain spatial jump degree. The average single charging power and the proportion of fast charging are used as energy preferences. The coefficient of variation of weekly charging frequency is calculated to obtain demand stability. Based on time regularity, spatial jump degree, energy preference and demand stability, a wide table of user-level behavior profiles is constructed to obtain the user behavior profile feature set.

[0019] In this embodiment, the method for obtaining the special user includes: The user behavior profile feature set is Z-score standardized. Based on the Bayesian information criterion, the optimal number of clusters in the Gaussian mixture model is determined by the elbow rule. The Gaussian mixture model is configured according to the optimal number of clusters. The standardized feature set is input into the Gaussian mixture model for training until the change in the log-likelihood function of adjacent iterations is less than or equal to the preset convergence threshold. The posterior probability matrix of each user sample belonging to each cluster is output. Based on the posterior probability matrix, each user sample is assigned to the cluster with the highest posterior probability to obtain the behavioral subgroup division results of each user. The information entropy of each user sample is calculated according to the posterior probability matrix as the soft attribution entropy, and the soft attribution entropy distribution of each user sample is obtained. The upper quartile and interquartile range are calculated according to the soft attribution entropy distribution using the box plot statistical method. Based on the upper quartile and interquartile range, the anomaly threshold is determined, and users whose soft attribution entropy exceeds the anomaly threshold are pre-marked as special users.

[0020] In this embodiment, the method for obtaining the conventional domain and the special domain includes: Based on the behavior subgroup segmentation results, the standardized user behavior profile feature sets within each behavior subgroup are input into the isolated forest for anomaly detection. Outlier scores of user samples within each subgroup are output. Outlier samples are filtered by time period according to the distribution of user outlier scores. Outlier samples are sorted in descending order according to outlier scores. Pre-labeled specific users and outlier samples are deduplicated and merged to obtain a set of specific samples. The remaining non-specific samples are used as regular samples. Based on the two types of samples, the prediction space is divided into a regular domain and a specific domain. The regular domain includes users and charging scenarios corresponding to regular samples, and the specific domain includes users and charging scenarios corresponding to specific samples.

[0021] In this embodiment, the method for obtaining the point prediction result includes: Based on the conventional domain, the charging demand for future periods is regressed and predicted using the first prediction model to obtain point prediction results, which are used as the baseline prediction values ​​for the conventional domain. For the prediction of abandonment points in the special domain, the charging demand deviation is obtained by aggregating the charging power difference between historical special periods and regular periods for each special user in the special domain. The sum of the baseline prediction value of the regular domain and the charging demand deviation is used as the prediction baseline for the special domain. The formula for calculating the charging demand deviation is as follows: ; in The amount of deviation in charging demand caused by special users. For time period Internal special user set, For user samples Historical anomalous order subsets User samples A subset of historical regular orders For orders The charging capacity, For orders The charging capacity.

[0022] In this embodiment, the method for obtaining the prediction confidence interval includes: Based on the specific domain, deviation sequences relative to the baseline of the conventional domain are extracted from historical demand during the same period. A tail sample subset is determined based on the deviation sequences. The upper bound of the interval is calculated based on the tail sample mean and tail sample standard deviation. The formula for calculating the upper bound of the interval is as follows: ; ; in For deviations from the sequence exceeding The tail sample subset of quantiles, As an off-sequence, For the first deviated sequence Demand deviation during a historically unique period , where is the number of tail samples. The upper bound of the interval is... This is the tail magnification factor. The summation index within the tail sample subset; The lower bound of the interval is determined based on the quantiles of the deviation sequence, and the formula for calculating the lower bound of the interval is as follows: ; in The lower bound of the interval, For the off-sequence quantiles; Based on the specific domain prediction baseline, upper bound, and lower bound, quantile regression is performed using a second prediction model to obtain the prediction confidence interval.

[0023] In this embodiment, the method for obtaining the charging demand prediction information includes: The density of specific samples is determined based on the ratio of the number of specific samples to the total number of active users in the current period, and the baseline deviation is calculated based on the degree of deviation between the specific domain prediction baseline and the regular domain baseline prediction value. The specific sample density, prediction confidence interval width, and baseline deviation are weighted and fused to obtain the prediction confidence level. Based on the numerical distribution of the prediction confidence level, quantiles are performed to obtain high confidence, medium confidence, and low confidence levels. Differentiated charging demand prediction information is output according to the confidence level, point prediction results, and prediction confidence interval. The differentiated charging demand prediction information includes scheduling early warning information.

[0024] A second aspect of the present invention also provides a machine learning-based electric vehicle charging demand prediction system, comprising: Data acquisition module: used to acquire charging order data and construct a user behavior profile feature set based on historical charging order data. The profile feature set includes temporal regularity, spatial jump, energy preference, and demand stability. The module for dividing the regular and special domains is used to perform probabilistic soft clustering of user behavior profiles using a Gaussian mixture model, pre-label special users based on soft attribution entropy, filter outlier samples through an isolation forest, and divide the prediction space into a regular and special domains based on the preliminary labels and outlier samples. Point prediction module: used to make point predictions of charging demand in future periods based on the conventional domain, abandon point prediction for special domains, and output prediction confidence intervals based on historical deviation information of the same period in future periods. Demand forecasting information output module: used to evaluate the forecast credibility based on the point forecast results, forecast confidence interval and current specific sample density, and obtain differentiated charging demand forecasting information based on the forecast credibility.

[0025] The above description is merely an example and illustration of the structure of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the structure of the invention or exceed the scope defined in the claims, all of which should fall within the protection scope of the present invention.

Claims

1. A machine learning-based method for predicting electric vehicle charging demand, characterized in that, Includes the following steps: Obtain charging order data and construct a user behavior profile feature set based on historical charging order data. The profile feature set includes temporal regularity, spatial jump, energy preference, and demand stability. User behavior profiles are probabilistic soft clustering using a Gaussian mixture model. Specific users are pre-labeled based on soft attribution entropy, and outlier samples are filtered out using an isolation forest. The prediction space is divided into a regular domain and a specific domain based on the preliminary labeling and outlier samples. Based on the conventional domain, point predictions are made for charging demand in future periods, while point predictions are abandoned for special domains. Confidence intervals are output based on historical deviation information from the same period in future periods. The prediction reliability is assessed based on the point prediction results, prediction confidence interval, and current specific sample density. Based on the prediction reliability, differentiated charging demand prediction information is obtained.

2. The electric vehicle charging demand prediction method based on machine learning according to claim 1, characterized in that, The method for obtaining the user behavior profile feature set includes: Historical charging order data is anonymized, user-level aggregation is performed based on the anonymization identifier, and historical charging behavior statistics for each user are extracted through a preset sliding time window. The historical charging behavior statistics include the charging period, charging duration, charging station identifier, charging amount, and charging power for each order. The charging frequency distribution for each charging period is calculated based on the start and end times, and the information entropy and standard deviation of the charging frequency distribution for each period are used as the degree of time regularity. Geographical distances between charging stations are obtained based on charging station identifiers, and the number of charging stations is counted to obtain spatial jump degree. The average single charging power and the proportion of fast charging are used as energy preferences. The coefficient of variation of weekly charging frequency is calculated to obtain demand stability. Based on time regularity, spatial jump degree, energy preference and demand stability, a wide table of user-level behavior profiles is constructed to obtain the user behavior profile feature set.

3. The electric vehicle charging demand prediction method based on machine learning according to claim 1, characterized in that, The method for obtaining the special user includes: The user behavior profile feature set is Z-score standardized. Based on the Bayesian information criterion, the optimal number of clusters in the Gaussian mixture model is determined by the elbow rule. The Gaussian mixture model is configured according to the optimal number of clusters. The standardized feature set is input into the Gaussian mixture model for training until the change in the log-likelihood function of adjacent iterations is less than or equal to the preset convergence threshold. The posterior probability matrix of each user sample belonging to each cluster is output. Based on the posterior probability matrix, each user sample is assigned to the cluster with the highest posterior probability to obtain the behavioral subgroup division results of each user. The information entropy of each user sample is calculated according to the posterior probability matrix as the soft attribution entropy, and the soft attribution entropy distribution of each user sample is obtained. The upper quartile and interquartile range are calculated according to the soft attribution entropy distribution using the box plot statistical method. Based on the upper quartile and interquartile range, the anomaly threshold is determined, and users whose soft attribution entropy exceeds the anomaly threshold are pre-marked as special users.

4. The electric vehicle charging demand prediction method based on machine learning according to claim 1, characterized in that, The method for obtaining the regular domain and the special domain includes: Based on the behavior subgroup segmentation results, the standardized user behavior profile feature sets within each behavior subgroup are input into the isolated forest for anomaly detection. Outlier scores of user samples within each subgroup are output. Outlier samples are filtered by time period according to the distribution of user outlier scores. Outlier samples are sorted in descending order according to outlier scores. Pre-labeled specific users and outlier samples are deduplicated and merged to obtain a set of specific samples. The remaining non-specific samples are used as regular samples. Based on the two types of samples, the prediction space is divided into a regular domain and a specific domain. The regular domain includes users and charging scenarios corresponding to regular samples, and the specific domain includes users and charging scenarios corresponding to specific samples.

5. The electric vehicle charging demand prediction method based on machine learning according to claim 1, characterized in that, The method for obtaining the point prediction result includes: Based on the conventional domain, the charging demand for future periods is regressed and predicted using the first prediction model to obtain point prediction results, which are used as the baseline prediction values ​​for the conventional domain. For the prediction of abandonment points in the special domain, the charging demand deviation is obtained by aggregating the charging power difference between historical special periods and regular periods for each special user in the special domain. The sum of the baseline prediction value of the regular domain and the charging demand deviation is used as the prediction baseline for the special domain. The formula for calculating the charging demand deviation is as follows: ； in The amount of deviation in charging demand caused by special users. For time period Internal special user set, For user samples Historical anomalous order subsets User samples A subset of historical regular orders For orders The charging capacity, For orders The charging capacity.

6. The electric vehicle charging demand prediction method based on machine learning according to claim 1, characterized in that, The method for obtaining the predicted confidence interval includes: Based on the specific domain, deviation sequences relative to the baseline of the conventional domain are extracted from historical demand during the same period. A tail sample subset is determined based on the deviation sequences. The upper bound of the interval is calculated based on the tail sample mean and tail sample standard deviation. The formula for calculating the upper bound of the interval is as follows: ； ； in For deviations from the sequence exceeding The tail sample subset of quantiles, As an off-sequence, For the first deviated sequence Demand deviation during a historically unique period , where is the number of tail samples. The upper bound of the interval is... This is the tail magnification factor. The summation index within the tail sample subset; The lower bound of the interval is determined based on the quantiles of the deviation sequence, and the formula for calculating the lower bound of the interval is as follows: ； in The lower bound of the interval, For the off-sequence quantiles; Based on the specific domain prediction baseline, upper bound, and lower bound, quantile regression is performed using a second prediction model to obtain the prediction confidence interval.

7. The electric vehicle charging demand prediction method based on machine learning according to claim 1, characterized in that, The method for obtaining the charging demand forecast information includes: The density of specific samples is determined based on the ratio of the number of specific samples to the total number of active users in the current period, and the baseline deviation is calculated based on the degree of deviation between the specific domain prediction baseline and the regular domain baseline prediction value. The specific sample density, prediction confidence interval width, and baseline deviation are weighted and fused to obtain the prediction confidence level. Based on the numerical distribution of the prediction confidence level, quantiles are performed to obtain high confidence, medium confidence, and low confidence levels. Differentiated charging demand prediction information is output according to the confidence level, point prediction results, and prediction confidence interval. The differentiated charging demand prediction information includes scheduling early warning information.

8. A machine learning-based electric vehicle charging demand forecasting system, used to execute the machine learning-based electric vehicle charging demand forecasting method according to any one of claims 1 to 7, characterized in that, The system includes: Data acquisition module: used to acquire charging order data and construct a user behavior profile feature set based on historical charging order data. The profile feature set includes temporal regularity, spatial jump, energy preference, and demand stability. The module for dividing the regular and special domains is used to perform probabilistic soft clustering of user behavior profiles using a Gaussian mixture model, pre-label special users based on soft attribution entropy, filter outlier samples using an isolation forest, and divide the prediction space into a regular and special domains based on the preliminary labels and outlier samples. Point prediction module: used to make point predictions of charging demand in future periods based on the conventional domain, abandon point prediction for special domains, and output prediction confidence intervals based on historical deviation information of the same period in future periods. Demand forecasting information output module: used to evaluate the forecast credibility based on the point forecasting results, forecast confidence interval and current specific sample density, and obtain differentiated charging demand forecasting information based on the forecast credibility.

Images