Improved Modeling of Lithium-Ion Battery Capacity Degradation Using an Individual-State Training Method and Recurrent Softplus Neural Network

An individual-state training method using multiple battery cycling data and initial state training for individual batteries is proposed in this study. The training method is based on the Broyden-Fletcher-Goldfarb-Shanno quasi-Newton method, and is modified to adapt to different battery samples by training the initial states for individual batteries to improve the modeling precision. This is equivalent to training a different model for each battery with shared model parameters, which improves the generalization of the model while preserving the model’s capability to account for individual variation. Based on the relatively small sample size and experimentation, the simple recurrent neural network with softplus activation is used as the model. The full NASA battery degradation dataset containing 34 degradation tests under varied cycling conditions is used to validate the proposed method. Compared to methods without the individual-state training, the proposed method reduces the average prediction error under changing cycling conditions from over 1 Ah to under 0.2 Ah for a nominal capacity of around 2 Ah. This improvement in accuracy is attributed to using multiple cycling data during training and the initial state training for individual batteries.