SOH Estimation for Battery Energy Storage Systems Based on Multiscale Temporal Convolutional Network and Relaxation Voltage

The large-scale integration of renewable energy into power grids introduces significant challenges to stable operation of power systems due to its randomness and volatility. Battery energy storage systems (BESS) serve as a crucial solution for accommodating renewable energy and ensuring grid stability. Accurate state of health (SOH) estimation is crucial to ensure safe operation of BESS. However, the existing SOH estimation methods face challenges related to limited generalization in feature extraction and insufficient interpretability. In this paper, an SOH estimation method based on a multi-scale temporal convolutional network (MSTCN) with parameter-free attention and relaxation voltage is proposed. First, multi-category features are constructed from relaxation voltage data at three different scales: point features, statistical features, and cumulative degradation features, leveraging battery operating principles and degradation mechanisms. These features can be obtained from a short period at the initial stage of battery discharge, independent of a full charge–discharge cycle, thereby effectively isolating the influence of operating conditions on the estimation model. Subsequently, an MSTCN model is developed to map the relationship between these features and SOH. This model incorporates a parameter-free attention mechanism to adaptively focus on critical degradation information. Additionally, a comprehensive dataset consisting of degradation experiment conducted on 20 lithium-ion batteries under multiple charging-discharging conditions is used to validate the proposed method. The implementation results show that the constructed relaxation voltage degradation features exhibit Pearson correlation coefficients with SOH of no less than 0.9126, demonstrating effectiveness of the extracted degraded features. Furthermore, in cross-condition SOH estimation, MSTCN model achieved an average coefficient of determination (R²) of 0.9681, validating the effectiveness of the proposed algorithm.