Cycle based state of health estimation of lithium ion cells using deep learning architectures

State of Health estimation in lithium-ion batteries is critical for reliable operation in electric vehicles and energy storage systems. This work evaluates four deep learning models-Multilayer Perceptron, Gated Recurrent Unit, Long Short-Term Memory, and Temporal Convolutional Network for cycle-based SoH prediction using discharge data from the NASA B0005, B0006, and B0007 cells. SoH values were obtained by numerical integration of discharge current and normalized with respect to the initial capacity. All models were implemented in PyTorch and assessed using RMSE, MAE, and R² metrics. On B0005, the MLP achieved RMSE 0.0069, MAE 0.0049, and R² = 0.9955, with TCN showing similar accuracy. Results on B0006 and B0007 confirmed the stability of MLP and TCN predictions across different cells. Residuals remained tightly clustered, and loss curves indicated smooth convergence. GRU and LSTM required higher training time without accuracy improvements. MLP demonstrated the best balance of accuracy and computational efficiency, making it suitable for embedded battery management systems. TCN provided robust accuracy with moderate complexity. The results verify that data-driven deep learning methods can capture nonlinear degradation behavior consistently across multiple cells.