Comparative Study of Embedded Deep Learning Models for State of Charge Estimation of Lithium-Ion Batteries

Accurate state of charge (SOC) estimation is essential for the safe and efficient operation of lithium-ion battery systems, while deploying deep learning-based estimators in embedded battery management systems remains challenging due to limited memory, computational capability, and real-time constraints. To address this issue, this study comparatively investigates four deep learning models, namely CNN, LSTM, CNN–LSTM, and lightweight TinyCNN, for embedded-oriented SOC estimation. A unified framework integrating data preprocessing, model training, post-training INT8 quantization, and TensorFlow Lite Micro deployment on an ESP32 microcontroller is established. Results under the US06 driving condition show that all models can track the SOC trajectory, although their relative performance differs between FP32 and INT8 inference. CNN–LSTM achieves the lowest FP32 error, with an MAE of 1.31%, while LSTM achieves the lowest INT8 error, with an MAE of 1.70%. Although TinyCNN has a higher INT8 MAE of 2.55%, it contains only 2913 parameters, occupies 8.51 KiB in INT8 form, and achieves the lowest measured ESP32 mean inference latency of 41.992 ms. These results demonstrate the feasibility of the evaluated INT8 models on the ESP32 under the US06 condition and indicate that model selection should depend on application-specific trade-offs among estimation accuracy, model size, runtime memory, and inference latency.