HyBattNet: Hybrid Framework for Predicting the Remaining Useful Life of Lithium-Ion Batteries

Abstract

Accurate prediction of the Remaining Useful Life (RUL) of lithium-ion batteries is critical for ensuring the safety, longevity, and operational efficiency of electric vehicles and smart grid applications. However, battery degradation is a highly complex, non-linear electrochemical process. Existing data-driven models often struggle to track long-term health metrics because raw sensor signals—such as voltage, current, and capacity—are highly dynamic and prone to localized temporal fluctuations. This paper introduces HyBattNet, a sophisticated hybrid framework designed to precisely estimate remaining usable charge-discharge cycles. The system features an innovative dual-stage design: a delta-based preprocessing pipeline that denoises raw inputs and extracts statistical differences between successive cycles, followed by an advanced hybrid deep learning network. By merging continuous-time physics-based dynamics with traditional discrete sequence modeling, HyBattNet provides a highly robust, data-driven approach to battery health management. The HyBattNet framework advances battery prognostics by introducing a specialized signal preprocessing pipeline that computes derived capacity features and utilizes a delta-based method to effectively capture subtle degradation changes between consecutive charge-discharge cycles. At its core, the system deploys a unique hybrid deep learning architecture that integrates 1D Convolutional Neural Networks (CNN) for spatial feature extraction, Attentional Long Short-Term Memory (A-LSTM) to capture localized temporal dependencies, and Ordinary Differential Equation-based LSTM (ODE-LSTM) blocks to seamlessly blend continuous-time physical dynamics with discrete sequence-to-sequence modeling. Furthermore, the model incorporates transfer learning strategies that ensure high predictive stability and robustness even when fine-tuned on highly restricted target datasets. Rigorously validated on multiple public LFP/graphite lithium-ion battery datasets, HyBattNet demonstrates superior prognostic reliability, significantly outperforming conventional machine learning and standalone deep learning baselines by achieving a highly competitive Root Mean Squared Error (RMSE) of 101.59 cycles.