State-of-Health Estimation for Lithium-Ion Batteries Based on Initial Voltage Segmentation and Transfer Learning
Conventional state-of-health (SOH) estimation algorithms for lithium-ion batteries fail to extract requisite features when cells operate under random partial charge-discharge cycling, a condition prevalent in grid-scale energy storage. This study proposes an estimation framework predicated on segmenting the initial charging voltage. Capacity increment (IC) curves are analyzed to extract features corresponding to the initial charge voltage point. Random forest and a composite index determine the optimal feature set and cardinality, which subsequently define the segmentation intervals for the initial charging voltage. Within each interval, interval-specific features are employed for SOH estimation. To address the data scarcity that impedes model training for operational batteries, a transfer learning strategy is implemented. A sample-based transfer method, TrAdaBoost.R2, improved by dynamic time warping (DTW), estimates battery state. DTW computes similarity between source and target domain features, and this similarity is integrated into the weight update mechanism of TrAdaBoost.R2, enhancing convergence and computational speed while preserving accuracy. Validation against NASA and XJTU datasets demonstrates the method's efficacy. In simulation experiment 2, the improved TrAdaBoost.R2 achieves a root mean square error (RMSE) of 0.009, outperforming classical TrAdaBoost (0.022), Transfer Stacking (0.018), and Two-stage TrAdaBoost (0.021). The proposed approach offers a robust solution for SOH estimation under partial charging conditions with limited target-domain data.