Temporal convolutional network for prediction of Li-ion batteries capacity fade knee

Abstract

Lithium-ion batteries are widely used for electric vehicles for fast charging capabilities and driving range, but they are characterized by a deterioration dependent from their operational conditions. The process of battery degradation shows a first stage when the degradation proceeds at a slow pace, followed by a stage when the battery abruptly deteriorates. The transition between these two stages is known as capacity fade curve knee. Predicting the capacity fade curve knee can be used to improve the lifetime of the battery by modifying the charging strategies, planning the maintenance, deciding warranty conditions in more cost-effective way and is also important in second life (when the battery is used in a less-demanding application), by timely preventing an unreversible aging trend that could occur in the first life. The present study aims to develop a novel technique to predict the occurrence of the capacity fade curve knee, introducing an engineered state of health indicator and using a Temporal Convolutional Network (TNC) that combines dilations and residual connections with causal convolutions. The effect of the choice of different initial cycles on the prediction accuracy has been studied and a heuristic uncertainty quantification has been provided to obtain an approximate measure of the prediction quality. Moreover, the network has been trained using inputs with different sampling frequencies and results have been compared. The prediction metrics of the present method have been compared to other methods, showing the benefit of this method for the early forecast of the onset of the capacity fade curve knee.