Lightweight Deep Learning for SOC Estimation of Various Lithium-Ion Batteries on Xilinx ZCU104 FPGA

Accurate estimation of the State of Charge (SOC) in Lithium-ion batteries is critical for the safe and efficient operation of electric vehicles and energy storage systems. This paper proposes a lightweight deep learning framework for SOC estimation, targeting deployment on resource-constrained edge AI hardware. We evaluate several recurrent and convolutional neural network architectures using the UNIBO Powertools Dataset, with a GRU-based model achieving the best balance between accuracy and memory footprint. The GRU model, featuring a single recurrent layer followed by three dense layers, is reimplemented in C + + and deployed on a Xilinx ZCU104 FPGA. The deployment achieves a latency of 38.76 ms and an energy consumption of 59.70 mJ per inference, with minimal resource usage. These results demonstrate the suitability of our approach for real-time, low-power battery management applications and confirm the benefits of FPGA acceleration for embedded SOC estimation tasks.