A Multi-Source Time Series Fusion Model for Lithium-Ion Battery State-of-Health Prediction: An Empirical Study Based on Real-World Driving and Charging Data

Accurate prediction of battery State of Health (SoH) is critical for enhancing electric vehicles (EVs) performance and extending battery lifespan. However, existing approaches often rely on single-source data collected in laboratorial environments, failing to capture the dynamic impact of real-world driving and charging behaviors on battery degradation. This paper proposes a Transformer-based predictive framework that integrates multi-source temporal data from both driving behavior and charging processes. Leveraging real-world data from 263 EVs over one year, we design a "historical smoothing window" mechanism to model the long-term SoH evolution by mapping 7-day input segments to forecast the SoH values of 30 days later. By incorporating diverse driving features and electrical signal characteristics, the model employs a self-attention mechanism to capture long-range dependencies and inter-feature interactions. Experimental results demonstrate satisfying performance with an R<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> of 0.9484 and a MAPE of 0.7361% on the 30-day prediction task. This work provides a cost-effective and scalable SoH monitoring solution for industry applications and offers new insights for hybrid physics-informed and data-driven battery health modeling.