Analysis of EV User Charging Characteristics Based on Large Language Model Data Generation

With the increasing penetration of electric vehicles (EVs), users generate vast amounts of data during charging and driving processes. However, traditional data acquisition methods suffer from high collection costs, regional specificity, and privacy protection policies, resulting in low data utilization, difficulties in database construction, and insufficient support for fine-grained charging load analysis. To mitigate this problem, this study proposes a large language model (LLM)-based method for constructing an EV charging database. By efficiently fine-tuning the parameters of the Qwen2.5-7B model and integrating retrieval-augmented generation (RAG) technology, we develop the ChatEV.Data model for synthetic data generation. In simulation cases, we use the ChatEV.Data model to generate 500 multidimensional data samples based on different vehicle models, covering 19 key dimensions such as initial state of charge and charging power, thereby establishing a comprehensive EV database. This study validates the high-quality data generation performance of ChatEV.Data using the Euclidean Semantic Consistency (ESC) and the Data Rationality Index (DRI), demonstrating the authenticity of the generated data. Additionally, we compare our results with data produced by traditional probabilistic models. Through tail data edge detection, we prove that our database construction method enhances the identification of charging behaviors among marginal EV users.