Prediction of Battery Capacity Based on Deep Residual Network

Abstract

Consistency is essential to the life of battery packs. Therefore, there is a special process to determine the capacity of lithium batteries in their production process (aka grading). However, this process takes a very long time. We propose a new method based on deep learning, which uses data collected by sensors before the grading process to predict the battery capacity, hoping to reduce the time consumed in the whole process. We propose an end-to-end battery capacity prediction model. In our processing steps, complex feature extraction steps are not needed. On the contrary, we use a residual network to complete it automatically. We modified the original ResNet to suit our task. Convolution1D and global pooling layers are used to extract the time series feature. To improve the model's accuracy, we design a fusion model to deal with the time series of multi-step processes. Transfer learning is applied to help us train the model faster. The results on the test set show that the root mean square error of the predicted capacity of our fusion model is 4mAh, which is a 45% decline compared with the benchmark model. We visualize the extracted features, interpret the model and explain the possible mechanism of our model. Furthermore, based on our analysis, suggestions for improving prediction performance are put forward.