Adaptive multi-task learning for battery health prediction: A BiLSTM framework with Bayesian optimization and dynamic task weighting

Abstract

Accurate and reliable prediction of lithium-ion battery health is critical for extending battery life, ensuring operational safety, and optimizing energy management in electric vehicles and energy storage systems. However, battery degradation exhibits complex, nonlinear, and battery-specific patterns, which pose significant challenges to conventional single-task learning methods. This study proposes a Multi-Task Learning Bidirectional Long Short-Term Memory framework that concurrently predicts key battery health indicators, including cycle life, voltage decay rate, and temperature change rate. To enhance model efficiency and adaptability, we introduce an adaptive prediction window, optimized via Bayesian hyperparameter tuning, which dynamically adjusts based on the battery's degradation characteristics. Furthermore, a dynamic task weighting mechanism is incorporated to adjust task priorities based on a loss-driven strategy, thereby improving learning efficiency and prediction accuracy. Experimental validation using the NASA battery dataset demonstrates that the proposed framework significantly outperforms traditional machine learning models, yielding substantial improvements in the accuracy of all health indicators. The adaptive prediction window and dynamic task weighting mechanisms are pivotal in enhancing model generalization and robustness. These findings highlight the framework's potential for real-time battery health monitoring and predictive maintenance applications.