A Novel Hybrid Modeling Framework Integrating Feature Engineering for Battery Remaining Useful Life Prediction

Accurate remaining useful life (RUL) prediction is critical for the reliable operation of lithium-ion batteries. Traditional data-driven methods often suffer from parameter redundancy and error accumulation in state prediction. This paper proposes a hybrid data-driven RUL prediction framework based on Gaussian process regression (GPR) optimized by the lightning search algorithm (LSA). First, both local and global indirect health features (HFs) are extracted from the external characteristic parameter curves and the incremental capacity curves during battery charging/discharging. Second, the Pearson correlation coefficient is applied to select highly relevant features, forming a compact feature set. Third, a GPR model is developed, and the LSA is introduced to optimize its hyperparameters, overcoming the tendency of the conjugate gradient method to fall into local optima or fail to converge. Experimental results under identical conditions show that the proposed LSA–GPR model achieves a prediction error of 3% or less.