Robust State of Health Estimation for On-Road Electric Vehicles Using an LSTM-Improved iTransformer Hybrid Network

Accurate estimation of the State of Health (SOH) of lithium-ion batteries is essential for ensuring the safety, efficiency, and lifecycle management of electric vehicles (EVs). Although data-driven approaches have become the mainstream solution for SOH estimation, most existing studies rely heavily on laboratory datasets collected under controlled and idealized conditions. Such datasets fail to capture the stochastic characteristics of real-world vehicle operation, including fragmented charging behaviors, varying environmental conditions, and significant sensor noise. Moreover, single deep learning architectures often struggle to simultaneously model the long-term temporal evolution of battery degradation and the complex multivariate correlations among operational variables. To address these challenges, this study proposes a hybrid neural network framework termed Long Short-Term Memory (LSTM)-improved iTransformer, which integrates the temporal modeling capability of LSTM networks with the multivariate feature interaction ability of an improved inverted Transformer architecture. In addition, a high-fidelity dataset was constructed using operational data collected from ten real-world electric vehicles. To simulate a realistic cloud-deployment scenario, a strict cross-vehicle validation strategy was adopted, where data from seven vehicles were used for model training and data from three entirely unseen vehicles were reserved for testing. The experimental results demonstrate that the proposed framework significantly outperforms conventional baseline models. In the multi-vehicle experiment, the model achieved a root mean square error (RMSE) of 1.18% and a coefficient of determination (R2) of 0.97, indicating promising cross-vehicle prediction performance within the available fleet. Furthermore, in the single-vehicle robustness experiment with limited training data, the proposed model achieved the lowest prediction error with an RMSE of 0.0045% and an MAE of 0.0034%, demonstrating superior accuracy and robustness compared with baseline models. These results suggest that the proposed method is a promising solution for battery health monitoring based on real-world operational data.