Technical Evolution, Bottlenecks, and Prospects of Long Short-Term Memory Networks in State-of-Health Estimation of Lithium-Ion Batteries

Abstract

State-of-health (SOH) estimation and remaining useful life (RUL) prediction of lithium-ion batteries are core technologies for ensuring the safe and economical operation of energy storage systems. Long short-term memory (LSTM) networks, by virtue of their ability to model long-range dependencies in temporal degradation processes, have become a widely adopted family of deep learning algorithms in this domain. Departing from the traditional chronological review paradigm, this paper constructs a five-stage analytical framework of architectural evolution along the functional essence of model architectures: “standard LSTM -> bidirectional LSTM -> attention-enhanced LSTM -> hybrid architectures -> physics-informed fusion”. Hyperparameter optimization is discussed separately as a cross-cutting strategy applicable across these architectural stages rather than as an additional evolutionary stage. More importantly, rather than merely enumerating performance metrics, this work starts from the structural characteristics and validation limitations of LSTM-based models and highlights three critical issues: the challenge of tracking strongly nonlinear late-life degradation, the lack of cross-domain consistency evidence for attention weights, and the confounding between parameter inflation and architectural innovation in reported accuracy gains. On this basis, we propose priority research pathways including grey-box physics-informed modeling, asymmetric uncertainty quantification, and edge-side lightweight deployment.