Abstract Accurate prediction of the Remaining Useful Life (RUL) of lithium–ion batteries is a critical enabler for the safety, reliability, and energy efficiency of modern electric vehicles (EVs). However, the nonlinear, multi-scale, and condition-dependent nature of battery degradation presents formidable challenges for conventional prognostic models. This work proposes a high-performance hybrid prognostic architecture that synergistically integrates (i) multi-resolution feature extraction via the Discrete Wavelet Transform (DWT), (ii) long-range temporal dependency modeling through an encoder–decoder Transformer network with multi-head self-attention, and (iii) nonlinear residual correction using XGBoost. To ensure globally optimal hyperparameter configuration and robust convergence, the full pipeline is optimized using the Chaotic Billiards Optimizer (CBO), complemented by local refinement with the Adam optimizer. Experimental evaluations conducted on benchmark battery aging datasets from the National Aeronautics and Space Administration (NASA) and the Center for Advanced Life Cycle Engineering (CALCE) demonstrate that the proposed framework substantially outperforms state-of-the-art deep learning and ensemble baselines, including recurrent neural networks, convolutionalrecurrent hybrids, transformer-based models, and gradient-boosted decision trees. The proposed approach achieves performance improvements exceeding 15% in both mean absolute error and root mean square error, with an average prediction accuracy characterized by a mean absolute error below 0.020, a root mean square error below 0.032, and a coefficient of determination exceeding 0.98. Ablation analyses further confirm the complementary contributions of multi-scale signal decomposition, attention-based temporal modeling, residual learning, and chaotic meta-heuristic optimization. Despite its hybrid structure, the framework remains computationally efficient, converging within a limited number of training epochs and enabling real-time inference (approximately 0.038 seconds per prediction window) with a lightweight model size of 2.14 million parameters, highlighting its suitability for embedded battery management systems. Overall, the proposed framework establishes a robust and interpretable foundation for next-generation battery prognostics, enabling intelligent predictive maintenance, enhanced safety, and energy-aware management in electric mobility systems.
Mchara et al. (Sat,) studied this question.