Capacity estimation of batteries: Influence of training dataset size and diversity on data driven prognostic models

Prognostics of batteries involve state estimation and remaining useful life (RUL) prediction. Various data-driven approaches are being studied to achieve accurate RUL predictions and SOH estimations to ensure safety and reliability of battery systems. The Gaussian Process Regression (GPR) is a statistical approach that accommodates the nonlinear nature and small sample size data to effectively predict the RUL of lithium-ion batteries. Artificial Neural Networks (ANN) have the ability to approximate nonlinear data and since battery degradation is a nonlinear process, neural networks-based models can provide accurate RUL predictions for lithium-ion batteries. In this paper, both the GPR model and ANN model approaches are implemented on the NASA PCoE battery datasets and the predictions are compared. The model training and validation is performed by splitting the data into various proportions (30:70%, 50:50%, and 70:30% for training: validation). Furthermore, the model's prediction accuracy is compared with respect to single vs multi-variable data and single battery data vs combined multiple battery data. The approach improves the prediction accuracy with the models exhibiting low RMSE values of 0.0181 and 0.0367 for GPR and ANN models, respectively, for B0018 and RMSE values of 0.1516 and 0.0810 for GPR and ANN models, respectively, for B0048