This study investigates data-driven state-of-health forecasting for second-life lithium-ion battery modules operated under long-term residential microgrid duty cycles. While CNN–recurrent–attention models are widely used in laboratory SOH studies, reliable field deployment is limited by distribution shift and by evaluation leakage when overlapped time windows are used across train and test. We address these issues using a two-stage domain-adaptation framework: a CNN–BiLSTM multi-head attention model is first pre-trained on NASA laboratory aging data to learn general degradation representations, then selectively adapted to microgrid data by freezing the CNN feature extractor and the lower BiLSTM while fine-tuning the upper temporal and attention/regression layers. For microgrid scenarios, we enforce a strict chronological split using date-labelled quarter folders and construct sliding windows only after splitting, preventing raw-sample overlap between train/validation/test. Experiments show low error on NASA cells (average RMSE ≈ 1.24% and MAE ≈ 0.98%) and robust transfer performance on three microgrid scenarios with RMSE below 3.5% (PV: 2.76%, PV-TOU: 2.34%, TOU: 3.12%). These results indicate that selective freezing reduces target-data requirements while maintaining generalization under real duty-cycle variability, supporting practical SOH monitoring and maintenance planning for second-life battery energy storage systems.
Errakkas et al. (Thu,) studied this question.