Probabilistic lithium-ion battery state-of-health prediction using convolutional neural networks and Gaussian process regression

Abstract

As the world accelerates its renewable energy transition, lithium-ion batteries are increasingly being used for large scale energy storage. Due to the high cost of production and replacement of large lithium-ion battery packs, accurate estimation of battery state-of-health (SOH) has become important. Accurate estimation is especially important for microgrid and transportation industries where operational batteries will be charged intermittently and incompletely. Incomplete charging scenarios provide models with inconsistent inputs, hence, cause inconsistent output accuracies, which are important to quantify in the context of dispatch controllers and predictive maintenance. In this paper, we present a model that provides a probabilistic prediction of battery SOH from periodic charging events as would be encountered in real world operations. The proposed model combines the feature recognition and many-to-one mapping abilities of the classic convolutional neural network (CNN) with the probabilistic characteristics of Gaussian Process Regression (GPR) models. The combined CNN-GPR model predicts battery SOH to less than 1% mean absolute error using random partial charge segments as input and displays its confidence in the prediction using the variance from the GPR. Additionally, the CNN-GPR is able to predict battery SOH irrespective of battery usage history; this is demonstrated by validation on dynamic data sets where the model achieves a mean absolute error of less than 1% despite only being trained on standardized degradation cycles.