Cluster-aware and feature-guided deep learning framework with fusion weighting for state of health prediction of li-ion batteries

Abstract

Accurate prediction of state of health (SOH) is critical for ensuring reliability and safety in lithium-ion batteries. This paper proposes a new cluster-aware and feature-guided deep learning (CAFG-DL) framework with a fusion weighting strategy for accurate and robust prediction of SOH. The approach begins by extracting six critical health features (CHFs) from discharge cycles, capturing voltage, temperature, and incremental capacity-based degradation markers from the NASA battery dataset across multiple cells. Following correlation analysis, these features are used to cluster battery cycles using density-based spatial clustering of applications with noise (DBSCAN), which was chosen due to its ability to identify complex, non-convex degradation patterns after comparison with K-means and hierarchical methods. For each cluster, localized deep learning models, bidirectional long short-term memory, gated recurrent unit, and feedforward neural network are trained to model intra-cluster temporal dynamics. After the training phase, a multi-criteria fusion method that assigns weights to each model’s prediction is introduced and used during the testing phase. The framework captures complex temporal dependencies and adapts predictions based on spatial proximity, SOH similarity, and model confidence. Three case studies are selected based on the CHF’s impact on SOH for each proposed algorithm, and four performance indices are considered for comparison. The CAFG-BiLSTM consistently outperforms conventional LSTM and cluster-based baselines, achieving a minimum RMSE of 0.0025 and MAPE of 0.18%. The framework demonstrates superior adaptability to heterogeneous aging behaviors and provides a scalable, interpretable solution for real-world battery health monitoring applications.