In-situ battery life prognostics amid mixed operation conditions using physics-driven machine learning

Accurately predicting in-situ battery life is critical to evaluate the system's reliability and residual value. The high complexity of battery aging evolution under variable conditions makes it a great challenge. We extract 6 physical features from voltage relaxation data to indicate battery performance fading, and then use data-driven techniques to predict battery life without considering any usage information. The model performance is validated against a dataset of 74 cells involving three battery types under mixed operation conditions. Experimental results show that battery lives are predicted accurately with the root-mean-squared-errors and mean absolute percentage errors being, respectively, generally less than 60 cycles and 10%. And the battery lives are classified quickly with the accuracies larger than 90%. This high prediction accuracy is maintained when only 6 sampling points taking 3–12 min are used. This work highlights the promise of using physics-driven machine learning to predict the behavior of complex systems under variable conditions.