Resource-Constrained Embedded Neural Network SOC/SOH Estimation for Real-Time On-Device Battery Management
TL;DR
The proposed framework uses temporal and dilated convolutions with residual learning and dual estimation heads, followed by structured pruning, knowledge distillation, and INT8 quantization for lightweight deployment and provides a compact and scalable solution for on-device joint SOC–SOH estimation.
Abstract
Accurate and real-time estimation of battery State of Charge (SOC) and State of Health (SOH) is essential for reliable Battery Management Systems (BMS), particularly on resource-constrained embedded platforms. Existing neural-network-based approaches achieve promising estimation accuracy but often involve computationally intensive architectures that limit on-device deployment. To address this limitation, this study proposes a ResourceAware Dual-Head Temporal Convolutional Network (RA-DHTCN) for simultaneous SOC–SOH estimation. The proposed framework uses temporal and dilated convolutions with residual learning and dual estimation heads, followed by structured pruning, knowledge distillation, and INT8 quantization for lightweight deployment. The model is implemented using Python and PyTorch and evaluated using the Kaggle EV Battery Charging & Thermal Runaway Dataset, containing 500 timestamped observations. The proposed model achieves 0.70% MAE, 0.84% RMSE, 0.76% MAPE, and for SOH estimation. The optimized architecture is designed to reduce model size, memory consumption, and inference latency while retaining estimation accuracy. Compared with conventional deep-learning baselines, the framework targets an [X%] improvement in estimation accuracy and supports realtime embedded BMS operation. Overall, RA-DHTCN provides a compact and scalable solution for on-device joint SOC–SOH estimation.