Abstract

Accurate prediction of battery State-of-Health (SoH) is crucial for ensuring safety, reliability, and longevity in energy storage systems. This study evaluates the utilization of Reservoir Spiking Neural Networks (RSNN) for battery SoH prediction, focusing on how network configuration affects prediction performance. Various RSNN architectures are analyzed by varying reservoir neuron number, connection density, inhibitory ratio, time step window, and readout network configuration. The findings show that a simple RSNN structure is sufficient to achieve high prediction accuracy, provided that the network is carefully configured to produce diversity in the spike count patterns within the reservoir layer. For the SoH dataset used, the original input features alone do not yield sufficient diversity, highlighting the importance of structural tuning in the reservoir. In the optimized network model, the prediction accuracy achieved a lowest error of 0.029 ± 0.011 RMSE, demonstrating strong accuracy and generalization across different battery datasets, with an energy consumption of approximately 150 nJ. Furthermore, this work provides a framework for implementing RSNN models on embedded platforms, including neuromorphic devices such as Intel Loihi and SoC-FPGA platforms with customized hardware circuits. The results demonstrate the potential of RSNNs as lightweight, efficient, and hardware-friendly models for practical battery SoH prediction.