Research on state of health prediction model for lithium batteries based on actual diverse data

Abstract

The state of health (SOH) is a key parameter for fault diagnoses and safety early warnings in the life cycle of lithium batteries in electric vehicles. The SOH prediction model generally uses the experimental data from the same batch of batteries in the same environment. These data may cause “overfitting” to the model as the attenuation of lithium batteries varies depending on the batch and working condition, especially in actual use. And there is a risk of serious deviation in the prediction result if there is no true value of the model. This paper proposes a SOH prediction model that evaluates the prediction uncertainty using data from different batches of batteries under actual working conditions. It not only quantitatively evaluates the credibility of the prediction model in absence of true values, but also filtering training data to improve the model accuracy and avoid overfitting. The model produces evaluation uncertainty for the prediction result based on the Gaussian process regression (GPR) method. Experiments' results show that the evaluation uncertainty is better than the prediction variance of GPR. The accuracy of the prediction model using the minimum evaluation uncertainty as the training data screening is an order of magnitude higher than that using all data for training.