Battery State-of-Health Estimation based on Partial Discharge Curves

Abstract

Accurate state-of-health (SOH) estimation is essential for safe operation and reliable fault diagnosis of lithium-ion batteries. However, existing data-driven methods often rely on specific voltage windows or health indicators that may shift with aging, limiting their applicability when partial discharge events are common. This paper proposes a robust SOH estimation framework based on differential capacity (dQ/dV) curves for partial discharge conditions. The framework uses a reference-concatenated input representation that pairs current and initial-cycle dQ/dV curves, enabling the model to learn degradation-related electrochemical changes. To address incomplete discharge data, a history-based imputation strategy reconstructs missing curve segments using earlier cycles. An attention-based feedforward neural network (FNN-Attention) is developed to capture complex dependencies across voltage levels in the dQ/dV curves. Numerical studies are performed using the MIT-Stanford battery aging dataset, where the proposed FNN-Attention achieves 0.38% root mean square percentage error (RMSPE), an 11.6% improvement over the state-of-the-art reported in literature. With partial discharge curves retaining only 30% (70% of the dQ/dV curve is missing) of the full curve, the FNN-Attention maintains 0.47% RMSPE using history imputation and 0.89% without requiring historical data. The results demonstrate the effectiveness and practical applicability of the proposed framework for real-world battery management systems.