Electric Vehicle Battery State of Health Estimation Using LSTM: Comparing Centralized and Federated Approaches

Sammanfattning

Accurate estimation of battery State of Health is important for range planning, safety margins, warranty handling, and second life use. This work compares centralised and federated training of the same Long Short Term Memory model for State of Health prediction using the MIT battery degradation dataset. Time series of voltage, current, and temperature and other features are transformed into fixed length windows, and models are trained under a matched protocol. The split is done in the cell level to avoid leakage with an 80/10/10 split for train, validation, and test. Federated learning is configured with three clients, one per experimental batch, and synchronous weighted averaging. Under this matched design, the centralised model attains lower error in absolute terms, while the federated model reaches similar explained variance. On the held out test set, the centralised model achieves MAE = 0.0049 Ah, RMSE = 0.0091 Ah, and R² = 0.83. The federated model achieves MAE = 0.00622 Ah, RMSE = 0.01019 Ah, and R² = 0.82. The federated setting entails higher coordination cost, observed as additional communication payload and longer wall clock time per effective improvement. These results indicate that, for this dataset and split policy, federated training preserves most predictive performance relative to a centralised baseline while keeping raw data local. When data minimisation or locality requirements preclude central aggregation, federated learning offers a viable alternative with a modest accuracy gap and explicit communication and timing costs.