Improving state-of-health prediction of li-ion batteries: a hybrid semi-empirical gaussian process regression approach

Abstract

Accurate battery degradation modeling is a prerequisite for the cost-effective operation of Battery Energy Storage Systems (BESS). However, existing approaches often face a tradeoff between the extrapolative robustness of semi-empirical equations and the flexibility of data-driven methods. This paper proposes a physics-guided Hybrid Exponential-Gaussian Process Regression (GPR) framework to predict the capacity fade of Samsung INR21700- 50E lithium-ion cells. Leveraging a comprehensive dataset encompassing 71 distinct aging conditions, three modeling strategies are evaluated: a baseline Semi-Empirical Model, a Hybrid GPR, and Hybrid Exponential-GPR model. The results reveal that calendar aging exhibits complex, non-monotonic fluctuations with State of Charge (SoC) that traditional semi-empirical models fail to capture. While standard data-driven methods resolve this local nonlinearity, they are prone to physical inconsistencies, such as predicting capacity recovery in untrained regions. The proposed Hybrid Exponential-GPR overcomes these limitations by employing a composite kernel architecture—combining Dot-Product and Matérn kernels—to enforce physical consistency while retaining data-driven flexibility. The model statistically confirms the negligible impact of C-rate for this cell chemistry and demonstrates superior generalization capability, reducing the global Normalized Root Mean Square Error (nRMSE) from 14.78% (baseline) to 2.45%. These findings underscore the potential of physics-guided machine learning for reliable lifetime prediction in sparse, multi-dimensional operational spaces.