State of Charge Estimation of a Formula SAE Vehicle using Deep Learning Models

Accurate estimation of the State of Charge (SoC) in electric vehicle batteries is crucial for performance optimization, safety, and reliability - especially in high-demand applications such as electric racing. This work focuses on a battery SoC estimation scheme developed at the University of Udine as part of a student competition, where real-time energy management plays a key role. Traditional SoC estimation methods often suffer from noise and drift in voltage and current measurements, particularly under the rapid load changes typical of race conditions. To address these limitations, we propose a data-driven approach using deep learning, specifically a Recurrent Neural Network (RNN) with Long Short-Term Memory (LSTM) units. The proposed method utilizes instantaneous measurements of voltage, current, and charge to estimate consumption and compensate for signal inaccuracies. Multiple network architectures are evaluated, comparing single-layer LSTM models with cascaded configurations to identify the optimal balance between model complexity and performance. The results show that a two-layer cascaded LSTM architecture significantly improves estimation accuracy, achieving errors as low as 0.5% in most test scenarios. Given the constrained scope of the application - a specific battery pack during a race session - the model maintains a low computational footprint and requires only a simple training procedure. In conclusion, this study demonstrates that LSTM-based models offer a viable and efficient solution for real-time SoC estimation in electric vehicles, especially when applied to short-duration, high-performance use cases such as competitive racing.