Intelligent Architecture Design and Tuning of Deep Learning Model for Precise EV Battery State-of-Charge Estimation

Accurate state of charge (SoC) estimation is a fundamental challenge for battery safety, reliability, and range prediction in electric vehicles. This paper proposes a deep learning-based framework that addresses the limitations of manual architecture design by systematically optimizing a deep neural network (DNN) using Bayesian optimization. The proposed framework introduces a coordinated, data-driven optimization of both network architecture and training hyperparameters, including learning rate, layer depth, hidden units, training epochs, and batch size. This approach enables key architectural and training parameters to be optimally determined to identify an effective model configuration while maintaining computational efficiency suitable for practical battery management system implementation. The effectiveness of the approach is evaluated under multiple driving profiles, four ambient temperature conditions, and noisy environments to emulate realistic EV operation. Comparative evaluation is conducted against long short-term memory (LSTM), gated recurrent unit (GRU), and conventional model-based SoC estimation approaches under identical experimental conditions. The results demonstrate that the proposed framework achieves an average MAE of 1.547% and an average RMSE of 2.046%, outperforming LSTM and GRU in both accuracy and robustness. In addition, the framework maintains a lightweight structure and significantly reduces training and inference time, supporting its suitability for practical battery management system (BMS) implementation. These findings indicate that the proposed framework provides a robust and computationally efficient solution for real-time SoC monitoring in electric vehicles.