A physics-informed deep learning framework for remaining useful life prediction of lithium-ion batteries with feature subset construction

Addressing critical challenges in lithium-ion battery (LIBs) remaining useful life (RUL) prediction, namely weak feature correlations, limited model generalizability, and insufficient physical consistency, this study proposes a novel deep learning framework that integrates optimal feature subset construction and physics-informed constraints. To address inherent multi-scale variations and noise in raw health indicator sequences, Seasonal-Trend decomposition using Loess (STL) is employed, extracting structured temporal information from trend, seasonal, and residual components. This process enables multivariate feature analysis. Subsequently, a Wrapper-based feature selection algorithm identifies, reconstructs, and preserves the most predictive subset of health features, thereby strengthening feature relevance and capturing crucial degradation factors. For the model architecture, this paper introduces MSTEA-Net, which incorporates multi-scale temporal convolutional encoding and a cross-variable attention mechanism. This design comprehensively captures dynamic cross-feature dependencies and long-term temporal evolution patterns. Critically, to enhance physical plausibility and interpretability, a triple-composite loss function, integrating data-driven prediction errors with dual physics-based regularization terms, is formulated and applied during model training. Extensive experimental evaluations on the publicly accessible CALCE and TJU datasets demonstrate the superior performance of the proposed framework. It significantly outperforms mainstream benchmarks in both RUL prediction accuracy and end-of-life (EOL) cycle localization precision. Specifically, the method achieves reductions in MAE of 5.56% and 11.8% on the respective datasets, while consistently constraining EOL localization errors within a stringent threshold of two cycles.