A Novel Deep Learning Method for Real-Time Estimation of Lithium-Ion Battery Capacity

Accurate prediction of lithium-ion battery states in electric vehicles (EVs) is crucial for enhancing safety, economic efficiency, and environmental sustainability. However, current research on real-time, on-board estimation of battery states is limited. This paper aims to bridge this gap by introducing a method that employs an Auto-BiLSTM algorithm coupled with an innovative data preprocessing technique using a low pass filter (LPF) to accurately predict battery capacity. Initially, data from the NASA dataset were pre-processed with the LPF to mitigate measurement noise and smooth oscillations caused by capacity self-generation, thereby stabilizing the input for neural network processing. After optimizing the autoencoder settings, this refined data, and the encoder output were used to evaluate the most effective predictive model among four widely used architectures. The results indicate that an autoencoder with 50-dimensional hidden units paired with a BiLSTM model offers the highest accuracy and the smallest model size, making it suitable for hardware implementation. Finally, a comparative analysis with the latest deep learning-based prediction models demonstrates that our model achieves top-level accuracy currently available in the field. All the data and codes related to this work can be accessed at the GitHub repository: https://github.com/Kayaa1.