A lithium-ion battery remaining useful life prediction model based on multilayer perceptron expert networks and temporal feature composition

Abstract

Unscheduled downtime caused by lithium-ion battery failures in electric vehicles and energy storage systems poses a significant challenge for accurately predicting remaining useful life (RUL). Existing methods, however, typically depend on high-quality and comprehensive performance data, limiting their applicability in complex real-world scenarios. To overcome this limitation, we propose MECCA-Net, a novel neural network framework whose core component is a self-designed Temporal Pattern Composer (TPC) that adaptively captures multi-level and cross-scale temporal degradation patterns from limited discharge capacity data. MECCA-Net further integrates multi-layer denoising autoencoders, multi-head self-attention mechanisms, and a mixture-of-experts structure to enhance its generalization capability and robustness. The experimental results demonstrate that MECCA-Net reduces the Relative Error (RE) by approximately 40% on several authoritative lithium-ion battery lifespan datasets compared to the latest state-of-the-art models. Furthermore, this approach exhibits superior prediction accuracy and stability performance over mainstream time-series modeling techniques, showcasing its efficiency and practical value in lithium-ion battery health management and predictive maintenance. The source code and datasets are available at https://github.com/keepawakeyi/MECCA-NET.