Personalized charging recommendation method for user behavior large model

Abstract

The invention provides a personalized charging recommendation method of a BFM (Behaviral Foundation Model), and aims to improve the accuracy and personalized level of charging recommendation of an electric vehicle through multi-modal data fusion and real-time feedback optimization. The method comprises the steps of data collection and preprocessing, user behavior modeling, feature fusion, personalized recommendation model construction and reinforcement learning optimization. Behavior feature embedding is generated by using a large-scale pre-trained user behavior large model, and multi-dimensional features such as geographic information, electricity price fluctuation and charging station load are combined, so that personalized charging suggestions can be generated in real time, a recommendation strategy is continuously optimized through reinforcement learning, and closed-loop improvement is formed. Compared with the prior art, the dynamic change of user behaviors can be accurately captured, the comprehensiveness and real-time performance of recommendation are improved, and the method has remarkable application value.

Description

Technical Field

[0001] The present invention relates to the technical field of electric vehicle charging, specifically a personalized charging recommendation method based on a user behavior large model. This method realizes the personalized, dynamic optimization and precise scheduling of the charging recommendation system by combining large-scale pre-training and user behavior feature modeling, thereby effectively improving the charging efficiency of electric vehicles and the user experience. Background Art

[0002] With the increasingly severe global environmental problems, electric vehicles (EVs) have been widely promoted due to their environmental protection and low-carbon advantages, becoming an important development direction for future transportation. The popularization of electric vehicles has promoted the construction of charging infrastructure and put forward higher requirements for the intelligent level of the charging management system. As an important part of the electric vehicle charging service, the accuracy and personalization degree of the charging recommendation system directly affect the user's charging experience. However, there are still many deficiencies in the existing charging recommendation technologies, making it difficult to meet the growing personalized charging needs of users.

[0003] Currently, most charging recommendation systems still rely on rule-based methods or shallow machine learning models, usually using fixed rules or simple predictions based on historical data. Although these methods can provide certain charging suggestions, they cannot accurately capture the complex behavior patterns of users and are difficult to adapt to the dynamic changes of charging demands. The charging behavior of electric vehicle users is affected by multiple factors such as travel plans, charging prices, geographical locations, and charging station load status, while traditional static models are difficult to comprehensively consider these factors, resulting in low accuracy of the recommendation results. In addition, the existing recommendation systems mainly rely on static data analysis and lack the ability to dynamically respond to the real-time behavior of users and changes in the external environment. For example, factors such as the fluctuation of electricity prices, the real-time load situation of charging stations, and traffic conditions will affect the charging decisions of users, but the current recommendation methods cannot quickly adjust strategies to adapt to these changes, resulting in lagging or inaccurate recommendation results.

[0004] With the development of big data and artificial intelligence technologies, user charging behavior and its related environmental data can be regarded as multi-modal data, including historical charging records, vehicle information, geographical locations, dynamic electricity prices, traffic flows, etc. However, traditional charging recommendation methods have not fully utilized these multi-dimensional data for deep modeling, resulting in limited adaptability and accuracy of the recommendation system. The lack of personalized recommendation ability is also a major bottleneck in the current charging recommendation technology. There are obvious differences among different users in terms of charging time, location, price sensitivity, etc., while the existing recommendation systems often fail to fully explore the charging habits of individual users and are difficult to provide charging solutions that truly meet individual needs. For example, some users tend to charge at low electricity prices, while others are more concerned about the availability of charging piles. However, traditional methods are difficult to accurately capture these preferences, thus affecting the user experience.

[0005] In recent years, deep learning and large-scale pre-trained models have made remarkable progress in various intelligent recommendation fields, especially showing strong data learning and feature extraction capabilities in natural language processing and behavior modeling. However, in the field of charging recommendation, the application of deep learning and large models is still in the exploratory stage and has not been fully utilized. Traditional methods are difficult to learn the complex non-linear relationships of charging behaviors, while large-scale pre-trained models can automatically extract user behavior patterns in the context of big data and make more accurate charging recommendations by combining environmental factors. Therefore, how to use advanced artificial intelligence technologies to improve the intelligence level of charging recommendations, making them have stronger personalized adaptation capabilities and real-time response capabilities, has become an urgent problem to be solved in the current field of electric vehicle charging management. Summary of the Invention

[0006] Aiming at the deficiencies in the existing technology, the present invention provides a personalized charging recommendation method based on a user behavior large model (Behavioral Foundation Model, BFM). This method can overcome the static nature and limitations of traditional charging recommendation systems. Through large-scale pre-trained models and multi-modal data fusion, it realizes the personalized, real-time optimization and dynamic adjustment of charging recommendations, thereby improving the accuracy of charging recommendations and user satisfaction.

[0007] The specific technical solution is as follows:

[0008] A personalized charging recommendation method for a user behavior large model includes the following steps:

[0009] 1) Data collection and preprocessing: Collect users' charging behavior data (such as historical charging records, vehicle information), environmental data (such as electricity price fluctuations, charging station locations), and geographical information (such as the user's current location, geographical coordinates of the destination), and perform missing value filling, standardization processing, and feature extraction on the data. The formula is expressed as:

[0010]

[0011] In the formula, is the original data, is the mean of the data, is the standard deviation of the data, ensuring the standardization processing of the data.

[0012] Calculate the spatial distance between the user and the charging station. The formula is:

[0013]

[0014] In the formula, and are the user and the charging station respectively With the geographical coordinates, calculate the distance between the user and the charging station as a spatial feature.

[0015] 2) User behavior modeling: Based on the user's charging behavior data, construct a large user behavior model to generate behavior feature embeddings. Use a pre-trained large model for modeling and capture the dynamic changes in the user's charging behavior through time series modeling. The formula is expressed as:

[0016]

[0017] In the formula, is the charging behavior feature of the user at time , is the context feature embedding, reflecting the characteristics of the user's behavior changing over time.

[0018] 3) Multimodal feature fusion: Perform multimodal feature fusion on the user behavior features and environmental data (such as geographical information, electricity price changes, etc.). The resulting fused feature vector is expressed as:

[0019]

[0020] In the formula, is the user behavior feature embedding, is the geographical feature embedding, is the electricity price dynamic embedding, and the electricity price change trend is captured by using a time series model (such as LSTM) for the electricity price fluctuations.

[0021] 4) Construction of personalized recommendation model: Based on the fused features, construct a personalized charging recommendation model, train it through a multi-layer fully connected network (MLP), and generate the recommended charging station and charging time. The loss function is calculated through the recommendation accuracy loss and the charging time error , and the formula is expressed as:

[0022]

[0023] In the formula, is the real charging station, is the predicted probability; and the charging time error is:

[0024]

[0025] In the formula, is the real charging time, is the predicted charging time.

[0026] 5) Real-time feedback and optimization: By obtaining user feedback in real time (such as charging completion time, charging station selection, etc.), using reinforcement learning algorithms (such as deep Q-learning) to continuously adjust and optimize the recommendation strategy, thus realizing the closed-loop improvement of the charging recommendation system. The reward function of reinforcement learning is defined as:

[0027]

[0028] In the formula, is the current state, is the current action, is the reward, is the discount factor.

[0029] Preferably, the data preprocessing step includes:

[0030] Using Multiple Imputation to fill in missing values;

[0031] Normalizing all numerical features;

[0032] Extracting time series features for modeling the charging time pattern of users and calculating the spatial distance between users and charging stations.

[0033] Preferably, the user behavior modeling step uses the BERT or Transformer architecture, the pre-training task is Masked Token Prediction, and the fine-tuning task is to predict the next charging behavior, including charging time and charging location.

[0034] Preferably, the personalized recommendation model is trained using a multi-layer fully connected network (MLP), and the recommended charging stations and charging times are generated at the output layer. The loss function is calculated through the recommendation accuracy loss and the charging time error.

[0035] Compared with the prior art, the beneficial effects of the present invention are:

[0036] Through a large-scale pre-trained user behavior large model and multi-modal data fusion technology, the present invention can accurately capture users' charging habits and preferences, thereby providing more personalized and dynamic charging recommendations. Different from traditional static recommendation systems, the present invention can dynamically optimize the recommendation strategy in real time according to user behavior and environmental changes, thus improving the accuracy and timeliness of recommendations. Through the integration of multi-modal data, especially the fusion of electricity prices and geographical information, the recommendation system can consider various factors more comprehensively, improving the comprehensiveness and accuracy of recommendations. The real-time feedback optimization using deep reinforcement learning continuously improves the charging recommendation scheme, forming a closed-loop improvement, further enhancing the adaptability of the system and user satisfaction. In addition, the technical solution proposed by the present invention can reduce the charging waiting time, improve the charging efficiency, provide a better charging experience for electric vehicle users, and effectively improve the operation efficiency of charging stations. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following will briefly introduce the drawings required for the description of the specific embodiments or the prior art. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0038] Figure 1 It is a flowchart of the technical solution of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0039] To better illustrate the present invention, the technical solution of the present invention will be described in detail in combination with specific embodiments. The following content is the detailed steps and corresponding implementation details of the specific embodiments, aiming to help understand how the present invention solves the problems existing in the prior art and improve the intelligence and personalization level of the electric vehicle charging recommendation system.

[0040] S1. Data collection and preprocessing

[0041] In the present invention, data collection and preprocessing is the first step of the personalized charging recommendation method, aiming to provide accurate and effective input data for subsequent user behavior modeling and recommendation models. This step involves obtaining users' charging behavior data, environmental data, and geographical location information, and performing necessary cleaning, standardization, and feature extraction on these data to facilitate model training and optimization.

[0042] (1) Data collection

[0043] The goal of data collection is to obtain as comprehensively as possible the data that can reflect users' charging needs and charging environments, including but not limited to the following categories:

[0044] User behavior data: The user's historical charging records, including the time, location, charging duration, charging cost, etc. of each charging. These data can help the system identify the user's charging habits, such as the user's charging time period, charging frequency, and frequently visited charging stations, etc.

[0045] Vehicle information: Includes information such as the remaining battery power, vehicle model, charging requirements, etc. Different battery capacities and the current battery status (such as the remaining charging power) of electric vehicles will affect the charging time and charging method. Therefore, this information is of great significance to the recommendation system.

[0046] Charging station information: Includes information such as the location of the charging station, the number of charging piles, charging power, load status, etc. The equipment status, current load, and geographical location of each charging station directly affect the user's charging decision.

[0047] Environmental data: Includes real-time electricity price fluctuations, weather conditions, traffic flow, etc. The change of electricity price directly affects the user's charging choice, especially for price-sensitive users. Weather conditions may affect the user's travel plan, and traffic flow affects the availability of charging stations.

[0048] Geographical location information: Collect the geographical coordinates of the user's current location and destination, as well as the location coordinates of the charging station through GPS positioning technology. The distance between the user and the charging station is an important feature in the recommendation system, which can help the system infer the user's charging needs and preferences.

[0049] (2) Data preprocessing

[0050] After data collection, it is necessary to preprocess the data to ensure the quality and consistency of the data. The preprocessing steps include the following aspects:

[0051] Missing value imputation: In the actual data collection process, some data may be missing or incomplete, which will affect the effect of subsequent modeling. To solve this problem, the present invention adopts the multiple imputation method to impute missing values. Specifically, for each missing value, the model will estimate based on other known data to generate multiple imputation results to ensure the accuracy of the imputed data. The multiple imputation method avoids over-reliance on a single imputation result by considering the correlation between data.

[0052] Data standardization: The data of different features may have different dimensions or ranges. For example, the numerical ranges of electricity price and charging duration vary greatly. To ensure the consistency of the weights of different data in the model, the present invention performs standardization processing on numerical features. The standardization formula is:

[0053]

[0054] In the formula, is the original data, is the mean value of the data, is the standard deviation of the data. Through this standardization process, it is ensured that the data ranges of all features are within the same scale, avoiding the negative impact of the deviation between features on model training.

[0055] Feature extraction and construction: To improve the modeling accuracy of charging behavior, data preprocessing also includes feature extraction. Especially the extraction of time series features and spatial features:

[0056] Time series features: By extracting the time attributes of charging behavior data, the charging periodicity of users can be captured. For example, by extracting time features in dimensions such as hours, weeks, and months, the system can identify the charging preferences of users within specific time periods, such as the charging tendency during low electricity price periods at night.

[0057] Spatial features: The impact of geographical location on charging recommendations cannot be ignored. Spatial features are generated by calculating the distance between users and charging stations. The calculation formula is as follows:

[0058]

[0059] In the formula, is the user and the charging station the Euclidean distance between them, and are the longitude and latitude coordinates of the user and the charging station respectively. This calculation helps the system identify whether users choose nearby charging stations and consider the accessibility of charging stations.

[0060] Data cleaning: During the data preprocessing process, there may be some noisy data or outliers, which have an adverse effect on model training. Therefore, the present invention cleans the data and removes abnormal data that does not conform to the rules, such as data points with too short charging time and unreasonable load status.

[0061] Data formatting and label generation: All collected data needs to be converted into a format acceptable to the model. For example, converting the charging behavior data of users into time series data, and encoding features such as geographical information, user preferences, and charging duration into numerical data. At the same time, corresponding labels are generated. For example, predicting the target (such as predicting the charging demand of users within a certain time period) based on the historical charging records of users.

[0062] Through the above data collection and preprocessing steps, the system can obtain high-quality and structured input data, laying a foundation for subsequent user behavior modeling, recommendation model training, and optimization. The accuracy and integrity of this step directly affect the performance of the charging recommendation system. Therefore, fine preprocessing and data cleaning are crucial for the present invention.

[0063] S2. User Behavior Modeling

[0064] In the present invention, user behavior modeling is the core part of the personalized charging recommendation method. This step uses deep learning techniques, especially pre-trained large models (such as BERT, Transformer) to model the user's charging behavior. By learning from the user's historical charging data, the model can automatically identify and capture the user's charging preferences, behavior patterns, and dynamic changes, providing accurate personalized features for subsequent charging recommendations.

[0065] (1) User Behavior Feature Extraction

[0066] The user's charging behavior involves multiple dimensions, including charging time, charging location, charging duration, cost, etc. To effectively model these behavior data, the present invention uses a large-scale pre-trained model (such as BERT or Transformer) to extract deep features from the original behavior data. The specific steps are as follows:

[0067] Input Feature Processing: Use the user's historical charging records as input features, specifically including charging time, charging location, charging duration, charging cost, etc. By encoding these original data, a preliminary feature vector of the user's behavior is generated. These feature vectors will then be input into the pre-trained model for further processing.

[0068] Pre-trained Model: Use models such as BERT or Transformer to train the user behavior data. The pre-trained model is trained through self-supervised learning, using a large amount of charging data to capture the potential patterns in the user's charging behavior. In this way, the model can understand the user's charging preferences at different times and locations, forming a deep representation of the user's behavior.

[0069] (2) Behavior Sequence Modeling

[0070] Since the user's charging behavior is a time series problem, that is, the user's charging choices and behaviors have time dependencies, the present invention captures the dynamic changes in these time series through behavior sequence modeling.

[0071] Model Architecture: Adopt the Transformer architecture, and learn the context feature embedding by inputting the user's charging behavior features at different time points. The specific formula is:

[0072]

[0073] In the formula, is the time point of the charging behavior characteristics, is the context feature embedding at the time point , is the time point 1's context feature. Through recursive learning, the model can capture the changes in the user's charging behavior over time and identify the user's short-term and long-term charging preferences.

[0074] Time-dependent modeling: Since the charging behavior has periodicity and long-term dependence, the model needs to consider the fluctuations of the charging behavior over time. For example, users may have different charging demands in different seasons, on different dates, or during specific time periods (such as at night when the electricity price is low). Through time series modeling, the model can effectively identify these periodic changes and improve the accuracy of recommendations.

[0075] In the process of user behavior modeling, the training and optimization of the model are crucial steps. To ensure the effectiveness and accuracy of the model, the present invention adopts the following training strategies:

[0076] Pretraining task: In the pre-training stage, the model is trained through the masked token prediction task. That is, by randomly masking some input features, the model learns how to predict the masked part through other known features. This method enables the model to learn the context relationship of the user's behavior and effectively capture the user's charging pattern.

[0077] Fine-tuning task: On the basis of the pre-trained model, a fine-tuning task is carried out to further optimize the model. The goal of the fine-tuning task is to predict the user's next charging behavior (such as charging time, location, etc.) according to the user's historical charging data. Through fine-tuning, the model can more precisely adapt to the charging needs of individual users and improve the personalization level of charging recommendations.

[0078] After pre-training and fine-tuning, the model can generate the behavior feature embedding of each user, denoted as , for the subsequent recommendation model. The behavior embedding can accurately reflect the user's charging preferences and habits. For example, some users may be used to charging at night, while others prefer to charge during the day on weekdays. The behavior feature embedding converts these preferences into numerical features that the model can understand and process.

[0079] The behavior feature embedding not only includes explicit features such as charging time and location, but also implicitly contains the user's charging demand pattern, such as charging frequency, charging amount, etc. Therefore, It is a highly compressed and comprehensive feature representation that can provide accurate personalized charging recommendations to the recommendation system.

[0080] S3. Multimodal Feature Fusion

[0081] In the present invention, multimodal feature fusion is one of the key steps of the personalized charging recommendation method. By fusing features from different data sources (such as user behavior, geographical information, and electricity prices, etc.), the comprehensiveness and accuracy of the recommendation system can be improved. Different from the features of a single data source, combining information from multiple dimensions can better reflect the user's needs and charging environment, and improve the personalization and real-time nature of the recommendation.

[0082] (1) Fusion of Geographical Features

[0083] Geographical information is an important factor in charging recommendation. When users choose a charging station, they usually give priority to the distance of the charging station. Therefore, the present invention uses a graph neural network (GNN) to model the geographical distribution of charging stations. By encoding the spatial relationship of charging stations, the model can better understand the impact of the distance between the user and the charging station on the charging decision during the feature fusion process.

[0084] Graph Neural Network (GNN) Modeling: Geographical Feature Embedding The generation process is as follows:

[0085]

[0086] In the formula, is the feature representation of the -th layer of the charging station , is the association weight between the charging station and other charging stations , is the set of charging stations adjacent to the charging station . In this way, the model can effectively capture the spatial relationship between charging stations and provide relevant information about the geographical location for the recommendation process.

[0087] (2) Dynamic Embedding of Electricity Prices

[0088] The change of electricity price has an important impact on the user's charging behavior. Especially for price-sensitive users, the charging tendency is stronger during the period with lower electricity prices. Therefore, modeling the real-time change of electricity price is crucial for accurately predicting the user's charging time and charging location.

[0089] Time series modeling: The dynamic changes in electricity prices usually exhibit obvious time series characteristics. In this invention, a Long Short-Term Memory (LSTM) network is used to model the trend of electricity price changes. The LSTM network can effectively capture the long-term dependencies of electricity prices and predict the trend of electricity prices in the future. Dynamic embedding of electricity prices Based on the learning of the electricity price time series, it reflects the impact of electricity price changes on charging behavior.

[0090] (3) Fusion of behavioral characteristics and environmental data

[0091] In addition to geographical characteristics and electricity price characteristics, the charging behavior characteristics of users and environmental data (such as the load status of charging stations, traffic flow, etc.) should also be considered. In this invention, the fusion of user behavior characteristics and environmental data is achieved by concatenating them into a unified feature vector.

[0092] Feature concatenation: In the feature fusion stage, the behavioral characteristics of users , geographical information characteristics and electricity price information characteristics are concatenated together to form a comprehensive feature vector , which is represented as:

[0093]

[0094] This fusion method unifies the feature information of multiple dimensions into one vector, avoiding information loss and enhancing the mutual correlation between features. This vector is then used as input for the personalized charging recommendation model for processing and prediction.

[0095] By fusing the characteristics of multiple data sources such as user behavior data, geographical information, and electricity price changes, this invention can generate more comprehensive and accurate recommendation features. Specifically:

[0096] Enhance the accuracy of the recommendation system: The fusion of multi-modal features enables the system to comprehensively consider multiple factors (such as charging station location, electricity price fluctuations, and user preferences), thereby improving the accuracy and real-time performance of charging recommendations.

[0097] Provide personalized recommendations: By fusing multi-dimensional data such as the behavioral characteristics, geographical information, and electricity price of each user, more personalized charging recommendations can be generated to meet the needs of different users in different charging scenarios.

[0098] After multi-modal feature fusion, the obtained comprehensive feature vector It is input into the subsequent personalized recommendation model. In the model, the fused features will help generate charging station recommendation results and charging time prediction results. In this way, the charging recommendation system can make optimal charging suggestions based on the user's historical behavior and real-time environmental changes.

[0099] S4. Construction of Personalized Recommendation Model

[0100] In the present invention, the construction of the personalized recommendation model is the core step to achieve high-precision charging recommendation. After fusing multi-modal features, the model can generate personalized charging suggestions based on information such as the user's behavior, geographical location, and electricity price changes. The recommendation model adopts deep learning methods, specifically including a multi-layer fully connected network (MLP) for feature processing and prediction. The following are the detailed steps of model construction:

[0101] Input Features is a comprehensive feature vector obtained through multi-modal feature fusion, which includes multi-dimensional information such as user behavior features, geographical information features, and electricity price dynamic features. The input vector has a dimension depending on the specific number of features, and each dimension represents an important factor affecting charging recommendations. For example, the behavior features may include the user's charging time and location preferences, the geographical information features include the distance between the user and the charging station, and the electricity price features reflect the trend of real-time electricity price fluctuations.

[0102] The recommendation model uses a multi-layer fully connected network (MLP) for feature processing and prediction. This network structure consists of multiple layers of neurons, and the output of each layer serves as the input of the next layer, thereby gradually extracting the deep information in the input features.

[0103] Input Layer: The input layer receives the feature vector after multi-modal feature fusion and passes it to the first layer of the network. The number of neurons in this layer is the same as the dimension of the input features.

[0104] Hidden Layers: The model contains multiple hidden layers, aiming to extract the deep information in the input features through non-linear transformations. The neurons in each hidden layer are processed by activation functions (such as ReLU), and the non-linear activation functions can help the model capture complex feature relationships. The number of neurons in the hidden layers can be adjusted according to the complexity requirements of the model.

[0105] Output Layer: The number of neurons in the output layer is set according to the results to be predicted. In the present invention, the two main tasks of the output layer are to generate the recommended charging station and the charging time , so the output layer contains two neurons, corresponding to the prediction of the charging station and the charging time respectively.

[0106] To optimize the performance of the model, a loss function is used to measure the difference between the prediction result and the true value. The loss function consists of two parts: recommendation accuracy loss and charging time error.

[0107] Recommendation accuracy loss: It is used to measure the accuracy of the recommended charging stations. Given the actual charging station and the predicted charging station , the recommendation accuracy loss can be expressed as:

[0108]

[0109] In the formula, is the label of the true charging station, is the probability value of the predicted charging station. The purpose of this loss function is to maximize the prediction probability of the correct charging station, thereby improving the accuracy of the recommendation.

[0110] Charging time error: It is used to measure the gap between the recommended charging time and the actual charging time. Given the true charging time and the predicted charging time , the charging time error can be expressed as:

[0111]

[0112] The purpose of this loss function is to minimize the gap between the recommended charging time and the actual charging time and improve the prediction accuracy of the charging time.

[0113] The final total loss function is the weighted sum of these two parts of losses:

[0114]

[0115] In the formula, and are the weighting coefficients, which are used to balance the weights of recommendation accuracy and charging time error.

[0116] The training process of the recommendation model is optimized through the backpropagation algorithm. First, the output of the model, that is, the predicted charging station and charging time, is calculated through forward propagation. Then, according to the difference between the output and the true value, the loss function is calculated and the parameters in the model are adjusted through the backpropagation algorithm. During the optimization process, a gradient descent method (such as the Adam optimizer) is used to minimize the loss function.

[0117] Through multiple rounds of training, the model can gradually learn and optimize the prediction ability of the charging station and charging time, thereby providing more accurate charging recommendations.

[0118] After training is completed, the personalized recommendation model can generate personalized charging station recommendations and charging time predictions in real time based on the user's current behavioral characteristics and environmental information. This recommendation not only considers the user's historical charging habits but also dynamically adjusts factors such as real-time electricity prices and the load situation of charging stations to ensure the accuracy and timeliness of the recommendation results.

[0119] S5. Real-time Feedback and Optimization

[0120] In the present invention, real-time feedback and optimization are key steps to improve the accuracy and adaptability of the charging recommendation system. By continuously receiving real-time feedback from users, the system can dynamically adjust the recommendation strategy, thereby continuously optimizing the charging recommendation in a changing environment. This process is achieved through reinforcement learning methods, especially by optimizing and updating the recommendation strategy through the Deep Q-Network (DQN) algorithm.

[0121] (1) Real-time Feedback Collection

[0122] During the charging process, the user's behavior and feedback are available in real time, especially the choices regarding charging station selection and charging time. The system continuously monitors and collects the following types of feedback information:

[0123] Charging Station Selection Feedback: Record the charging station finally selected by the user and compare it with the recommended charging station. If the user selects the charging station recommended by the system, it indicates that the recommendation is accurate; if the user selects other charging stations, the system can analyze the reasons and adjust the recommendation strategy.

[0124] Charging Time Feedback: Record the start and end times of the actual charging by the user and compare them with the recommended charging time. If there is a significant difference between the actual charging time and the recommended time, the system needs to make adjustments to more accurately predict the user's future charging time requirements.

[0125] User Satisfaction Feedback: The satisfaction of the recommendation can be evaluated by conducting a questionnaire survey or directly obtaining the feedback score of the user on the charging recommendation. If the user feedback indicates that the recommendation is inappropriate, it means that the recommendation model needs to be optimized.

[0126] (2) Reinforcement Learning and Strategy Optimization

[0127] Based on the real-time feedback information, the system uses reinforcement learning to adjust and optimize the recommendation strategy. The reinforcement learning model guides the recommendation system to optimize in the correct direction through a reward mechanism. This process uses the Deep Q-Network (DQN) algorithm, and the specific steps are as follows:

[0128] State Definition: The state of the system It includes the real-time features of the user (such as charging demand, current location, load status of the target charging station, etc.) and environmental information (such as electricity price, traffic flow, etc.). Through these status information, the system can comprehensively evaluate the user's current charging demand and environment.

[0129] Action definition: Action refers to the recommended charging station and charging time Based on the input of the current state, the system decides the charging station and time to recommend to the user.

[0130] Reward function: Reward function It is used to measure the effectiveness of the system's recommendation strategy and reflect the quality of the recommendation result. The reward function consists of two main parts:

[0131] User satisfaction reward: By comparing the difference between the charging station actually selected by the user and the recommended charging station, the user satisfaction reward The calculation formula is:

[0132]

[0133] Wherein, is the actual charging time, is the recommended charging time. This reward item encourages the system to predict the charging time and location accurately.

[0134] System benefit reward: Evaluate the benefit of the system according to the balance of the charging station load and the operating cost. The system benefit reward is expressed as:

[0135]

[0136] Wherein, represents the grid load balance, is the operating cost. This reward item encourages the system to optimize the recommendation scheme, avoid grid load overload and reduce the operating cost.

[0137] (3) Policy update

[0138] Through deep Q-learning (DQN), the system continuously updates its recommendation strategy. After each feedback, the system updates the Q-value function according to the current state and the action taken , as well as the corresponding reward :

[0139]

[0140] Wherein, is the discount factor, representing the influence degree of future rewards. By continuously updating the value, the system can learn the optimal recommendation strategy, making the recommended charging stations and time more accurate, thus improving the user's charging experience.

[0141] As the charging recommendation process progresses, the system continuously optimizes the recommendation strategy based on real-time feedback. Each optimization adjusts the weights of the recommendation model through deep Q-learning, gradually improving the accuracy of the recommendation results. Through this dynamic optimization method, the system can adjust the recommendation strategy in real time under the changing environments such as grid load, electricity price, user behavior, and charging station status, ensuring the efficiency and accuracy of charging recommendations.

[0142] The present invention provides a method for personalized charging recommendation of a large user behavior model, which combines large-scale pre-trained models and multi-modal data fusion technologies, effectively improving the accuracy and personalization level of electric vehicle charging recommendations. By deeply modeling the user's charging behavior and integrating multi-dimensional features such as geographical location, electricity price dynamics, and charging station load, this method can generate optimal charging suggestions in real time. In addition, by introducing real-time feedback and reinforcement learning mechanisms, the recommendation strategy can be dynamically optimized to ensure that the system can still operate efficiently under different environments and changing user needs. This method not only improves charging efficiency and user satisfaction, but also effectively reduces the grid load pressure and operating costs, having broad application prospects and significant economic benefits.

Claims

1. A personalized charging recommendation method for a large user behavior model, characterized in that the method comprises the following steps: Collect multi-modal data related to user charging behavior, including user historical charging records (time, location, cost), vehicle battery information, charging station location information, electricity price fluctuation data, and traffic flow data; Preprocess the collected data, including missing value filling, data standardization processing, and feature extraction; Extract time series features and spatial features, and the spatial features are modeled by calculating the geographical distance between the user and the charging station; Based on the user charging behavior data, build a large user behavior model, generate behavior feature embeddings, and use the pre-trained large model to perform behavior sequence modeling; Combine user behavior features with environmental data (including geographical information, electricity price changes, etc.) for multi-modal feature fusion to generate charging recommendation features; Build a personalized charging recommendation model, recommend charging stations according to the generated features, and give a prediction of the charging time; According to the user's real-time feedback, use reinforcement learning to optimize the recommendation strategy and realize the dynamic adjustment of the recommendation strategy.

2. The personalized charging recommendation method for the user behavior large model according to claim 1, characterized in that, The data preprocessing step includes: Perform multiple imputation processing on the missing values in the user behavior data; Perform normalization processing on the numerical features; Extract time series features and spatial features.

3. The personalized charging recommendation method for a large user behavior model according to claim 1, characterized in that the large user behavior model adopts a BERT or Transformer architecture, is pre-trained based on the charging behavior data, and is fine-tuned through a masked prediction task to predict the next user charging behavior.

4. The personalized charging recommendation method for a large user behavior model according to claim 1, characterized in that the multi-modal feature fusion step models the charging station location distribution through a graph neural network (GNN), and uses an LSTM model to model the dynamic change of the electricity price.

5. The personalized charging recommendation method for a large user behavior model according to claim 1, characterized in that the personalized charging recommendation model is trained using a multi-layer fully connected network (MLP), and the recommended charging station and charging time are generated at the output layer.

6. The personalized charging recommendation method for a large user behavior model according to claim 1, characterized in that the reinforcement learning optimization step updates the recommendation strategy according to the user's real-time feedback through a deep Q-learning (DQN) algorithm.

Images