Multi-task neural network combined with adaptive dual Kalman filter for joint estimation of battery state-of-charge and state-of-health
Junting Bao, Yuan Mao, Weijie Zhang, Zhiming Zhang, Youbing Zhang
Abstract
Precise estimation of lithium-ion battery (LIB) state constitutes fundamental technology for battery management systems (BMSs), ensuring safe electric vehicles (EVs). To address accuracy limitations of traditional independent battery state estimation methods, this work proposes a joint state of charge (SOC) and state of health (SOH) estimation framework using dual-extended Kalman filter (DEKF) and multi-task neural networks. Firstly, the coyote optimization algorithm (COA) integrates with innovation matrix adaptive optimization for DEKF noise matrix, achieving high-precision joint parameter and SOC tracking on short-time scales. Secondly, image features are extracted from noise-reduced incremental capacity (IC) curves. Leveraging both time-series data and these image features on long-time scales, a Multi-task SOH Forecasting Neural Network (MTSOHFNN) is designed to enhance SOH prediction generalization capability. Finally, A novel mutual compensation mechanism implements SOC output identification results as SOH prediction inputs, while SOH capacity predictions dynamically update SOC estimation components. Validation using NASA battery aging dataset demonstrates SOC estimation MAE consistently below 3.5 %. For SOH estimation, the RMSE reaches as low as 0.33 % and remains below 0.653 % across all cycles, with all prediction accuracies significantly surpassing baseline methods.