Advancing Onboard SOC Estimation: Transfer-Learning-Enabled Transformer Models Trained on Stock EV CAN Datasets

Estimating the state of charge (SOC) is a wellstudied problem due to its critical role in determining the remaining battery energy and power, attracting extensive research. Machine learning and deep learning have become increasingly popular for addressing this issue as more battery data becomes available. On-board data is typically stored in time-series format through CAN communication. To enable real-time updates of pre-trained models, using the same data format-CAN datais an important consideration. In this paper, a CAN dataset is utilized for model development. This study proposes two models, A and B, based on the transformer architecture to solve the sequence-to-scalar time-series problem of predicting SOC in electric vehicles. Model A uses a simpler architecture, while Model B incorporates the complexity of a full transformer model. A comprehensive hyperparameter study is conducted for the superior Model B. Model B outperforms a non-transformer baseline under equal training resource conditions. The paper also reports on the preparation and preliminary testing of an experimental setup to implement these models in real time. Both Models $\mathbf{A}$ and $\mathbf{B}$ achieve convergence with minimal tuning, highlighting another advantage of this approach. While this research focuses on electric vehicles such as Tesla and Chevy Bolt, the results are extendable to other EVs, PHEVs, and HEVs.