Multi-task learning and voltage reconstruction-based battery degradation prediction under variable operating conditions of energy storage applications

Accurate prediction of state-of-health (SOH) degradation is critical for the intelligent management of lithium-ion batteries in energy storage systems (ESSs). However, variable operating conditions and long-term operation complicate the degradation prediction and impact the prediction accuracy. To tackle these challenges, this paper proposes multi-task learning with regularization (MTL-RL)-based degradation prediction approach through the incorporation of multi-attribute feature extrapolation and charging voltage reconstruction considering the variable operating conditions of ESSs. First, stable charging voltages are reconstructed based on internal resistance compensation, which addresses the significant voltage distribution differences caused by variations in current levels of ESSs. Then, the mechanism model and the reconstructed capacity-voltage curves are utilized to extract multi-attribute features to improve the prediction accuracy while considering multiple critical factors. Finally, an adaptive MTL-RL framework is established to predict SOH degradation by recursively extrapolated features with the consideration of long-term regularization, which reduces the input data requirement and improves the long-term prediction stability. The proposed approach is verified by the aging experimental data that simulates the charging/discharging of ESSs in peak shaving and valley filling. Compared with prevailing methods, the proposed method achieved higher prediction accuracy for the energy storage application.