Stacked temporal deep learning for early-stage degradation forecasting in lithium-metal batteries

Abstract The integration of artificial intelligence (AI) into energy storage prognostics presents a transformative approach for enhancing the safety, reliability, and longevity of next-generation battery technologies. This study introduces a robust temporal deep learning framework for predictive modeling of lithium-metal battery (LMB) degradation, with a focus on AI-driven health forecasting. A comprehensive dataset comprising 23 LMB cells—diverse in capacity, chemistry, and cycling conditions—was curated to train and validate a suite of sequential models including Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), Transformer networks, and multiple fully connected Deep Neural Networks (DNNs). These were subsequently integrated into a stacked ensemble meta-model (S-DNN) using an Extreme Learning Machine (ELM), designed to enhance forecast accuracy and generalization. The ensemble achieved superior performance with an RMSE of 0.026 Ah, R 2 of 0.9917, and CVRMSE of 0.6955%, outperforming all individual models. Crucially, the framework demonstrated strong early-stage prediction capabilities using only 15% of the cycling data, maintaining a CVRMSE below 6.5%. Rich regression analyses and error visualizations were used to support interpretability and deployment readiness. Limitations related to uniform temperature cycling and the need for broader cross-domain validation are acknowledged as directions for future work. This work advances the frontier of AI for prognostics by introducing an interpretable, generalizable, and ensemble-based architecture for real-time health monitoring in complex electrochemical systems.