Hierarchical Multimodal Large Language Model-Based Interpretation and Aggregation of PDP and ICE Explanations for Battery State-of-Charge Prediction in Electric Vehicles

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Accurate prediction of the state-of-charge (SOC) of batteries is essential for reliable energy management and safety in electric vehicles (EVs). Although machine learning (ML) models have shown strong predictive performance for SOC estimation, their practical development is hindered by limited interpretability, particularly in understanding how individual input variables affect both global model behavior and instance-level responses. Partial dependence plots (PDPs) and individual conditional expectation (ICE) plots are widely adopted visual explanation techniques. However, their interpretation typically relies on manual and fragmented analysis, requiring substantial expert involvement. This study proposes a hierarchical multimodal large language model (MLLM)-based framework that automatically interprets and aggregates PDP and ICE explanations for battery SOC prediction in EVs. Multiple specialized MLLM agents generate structured textual interpretations of PDP and ICE plots for each input variable, capturing global trends and individual-level heterogeneity. These interpretations are subsequently synthesized by a feature-level aggregation agent, which resolves inconsistencies and identifies dominant behavior patterns. A global aggregation agent consolidates the feature-wise explanations into a coherent system-level report summarizing key factors, interaction characteristics, and uncertainty considerations relevant to SOC prediction. The proposed framework provides a scalable and reproducible solution for transforming visual explainable artificial intelligence outputs into structured and decision-oriented knowledge, thereby enhancing the transparency and trustworthiness of ML-based battery management systems.