A two-stage deep learning framework for early-stage lifetime prediction for lithium-ion batteries with consideration of features from multiple cycles

Lithium-ion batteries are a crucial element in the electrification and adoption of renewable energy. Accurately predicting the lifetime of batteries with early-stage data is critical to facilitating battery research, production, and deployment. But this problem remains challenging because batteries are complex, nonlinear systems, and data acquired at the early-stage exhibit a weak correlation with battery lifetime. In this paper, instead of building features from specific cycles, we extract features from multiple cycles to form a time series dataset. Then the time series data is compressed with a GRU-based autoencoder to reduce feature dimensionality and eliminate the time domain. Further, different regression models are trained and tested with a feature selection method. The elastic model provides a test RMSE of 187.99 cycles and a test MAPE of 10.14%. Compared with the state-of-art early-stage lifetime prediction model, the proposed framework can lower the test RMSE by 10.22% and reduce the test MAPE by 28.44%.

Automated summary

This paper proposes a battery lifetime prediction method based on time series data, which accurately predicts battery lifetime by extracting features from multiple cycles and compressing data using an autoencoder. Compared with existing models, this method can effectively reduce prediction errors.