Volume 31: Clean Energy Technologies towards Carbon Neutrality

Unsupervised Extraction of Degradation Related Features from Battery Cycling Data via a Conditional Temporal Convolutional Autoencoder Jianxiang Wang, Shijie Weng, Xuesong Feng, Xiaoli Peng, Yong Xiang

Abstract

Clean energy production is often accompanied by battery storage systems that undergo complex degradation processes. Incremental capacity and differential voltage peaks are traditionally used for degradation analysis, but are sensitive to current rate, deep degradation, and battery chemistry. To enhance the robustness of degradation feature tracking, this study proposes a conditional temporal convolutional autoencoder. The unsupervised nature of the proposed method enables the extraction of degradation related features from data that have no peak features. Results show that the proposed method can extract features and reconstruct battery cycling curves with high fidelity. Furthermore, the extracted features are highly correlated with peak locations of incremental capacity and differential voltage curves. The extracted features also have less noise and no missing values compared to the peak locations. Prediction of peak locations from single encoding achieves mean absolute errors of 0.019 V and 2.4% state-of-charge. The proposed method is therefore potentially useful for battery degradation analysis and health assessment.

Keywords Lithium-ion battery, differential voltage analysis, incremental capacity analysis, neural network, unsupervised learning