LLM and Bayesian Network Integrated Simulation Framework for Electric Vehicle User Charging Behaviorals

Electric vehicle (EV) users’ behaviors are influenced by users’ willingness, which is not directly observable. Emerging large language models (LLMs) have advantages in handling this problem. However, existing studies usually use LLM to directly output EV users’ behaviors, which is limited by the LLM’s inherent issues: “randomness” and “hallucination”. To address these issues, this study proposes an LLM-Bayesian network (BN) integrated simulation framework. First, LLM is used to extract and label EV users’ willingness. Second, LLM generates an explicit causal structure and prior probability, which constructs a BN. Finally, Bayesian inference updates the BN’s prior probability. The BN outputs EV users’ charging behaviors. Experimental results show that in few-shot scenarios, LLM-BN achieves a 20.3% to 27.5% improvement in predictive accuracy compared to pure LLM methods. Moreover, LLM-BN addresses “randomness” and “hallucination” issues, and effectively suppresses violations of physical constraints and the generation of fake information. In practical applications, peak shaving and valley filling are achieved through electricity price adjustments, reducing peak demand by 25% and increasing off-peak demand by 30%.