Physics-Informed Bayesian Neural Network for Li-Ion Battery Continual Learning

In this work, a continual learning problem for time-varying Li-ion battery systems is investigated based on a Physics-Informed Bayesian Neural Network (PI-BNN) approach. To build an interpretable dynamic model, a second-order battery equivalent circuit model is incorporated into a BNN, and a pretrained multi-layer perception approximating the battery nonlinear function is connected between the hidden and output layers of the PI-BNN. The network parameters of the PI-BNN are designed as the unknown resistance and capacitance variables of the Li-ion battery to learn probability distributions, and then each uncertain parameter is updated online using new measurement data, where the learning rate is proportional to the variance of the parameter probability distribution. By considering the uncertainty in the evolution of the posterior distribution, the uncertain model parameter is learned with improved model adaptability, and the output voltage is predicted based on the PI-BNN model during the lifelong battery monitoring. The effectiveness of the proposed approach is validated through cyclic charge and discharge simulations, which achieves accurate dynamic Li-ion battery modeling and output voltage prediction compared with a basic physics-informed neural network.