A multi-scale temporal fusion network for accurate and uncertainty-aware battery remaining useful life prediction

MSTFNet synergistically fuses SCNN local patterns, Mamba global dynamics and Informer focusing under a Bayesian head to deliver accurate, calibrated battery‑RUL predictions with credible intervals for risk‑aware BMS decisions. Accurate prediction of the remaining useful life (RUL) of lithium-ion batteries requires the concurrent satisfaction of three technical demands: efficient modeling of long sequences, fusion of multi-scale features, and reliable quantification of prediction uncertainty. This study proposes MSTFNet, a multi-scale temporal fusion network that integrates a spatial convolutional neural network (SCNN) for local patterns, a Mamba state-space module for linear-time modeling of global dependencies, and an Informer module for sparse temporal focusing, all under a Bayesian inference head optimized via a combined negative log-likelihood and mean-squared-error objective. The Bayesian head not only calibrates uncertainty but also regulates multi-module feature fusion, achieving a cooperative synergy unavailable to single modules alone. Across three benchmark datasets (NASA, CALCE, and HUST), MSTFNet attains a superior balance between prediction accuracy and model efficiency, achieving up to 29.4 % lower RMSE than state-of-the-art baselines. Further analyses confirm well-calibrated predictive intervals, robust cross-dataset generalization under train-on-A/test-on-B and few-shot settings, and deployment feasibility in terms of latency, memory, and INT8/pruning performance. Ablation results substantiate that BNN-regulated joint optimization effectively enhances Informer collaboration, validating the proposed regulatory mechanism.