Optimization of Temporal Convolutional Network for Remaining Useful Life Prediction of LithiumIon Battery Using Separable Convolution

Predicting the remaining useful life (RUL) of lith-ium-ion batteries is crucial in battery management and mainte-nance for electric vehicles (EVs) and industrial machinery [1]. In this prediction, a data-driven approach using deep learning, which has demonstrated high performance in image recognition and natural language processing, has been developed, but there are still some issues to be solved. In particular, these include the increase in computational complexity in the hidden layers when introducing deep learning [3], as well as the improvement of accuracy and the practicality of attaching deep learning to battery-powered devices [3]. Therefore, in this study, we will solve the above problems by introducing a temporal convolutional network (TCN), which has numerous advantages for time-series data, and by devising the configuration of the network. First, we propose using depthwise separable convolution layers to reduce computational complexity and the number of parameters. These layers have been used in mobile application networks in recent years. In addition, we further improve computational efficiency by adopting bottleneck architecture centered on the pointwise convolution layers for the residual blocks in TCN. In this study, we performed RUL prediction for these networks using the NASA Ames Prognostics Centers of Excellence (PCOE) dataset and compared the models' accuracies and complexities using mean square error and other evaluation metrics. Therefore, while there was no degradation in the models' accuracies, the computational complexity and number of parameters were reduced by a factor of 10 using the separable convolution layers, and by a factor of 20 using the bottleneck architecture.