Data-Driven Timestep Analysis for State of Health Estimation Using BiLSTM Neural Networks

Precise assessment of battery State of Health (SOH) is fundamental for guaranteeing the reliability and longevity of energy storage systems. This study examines the impact of input timestep selection on a Neural Network (NN) model. In this study, a Bidirectional Long Short-Term Memory (BiLSTM) network is used to analyze how different input timesteps influence prediction accuracy. The model is trained on a battery degradation dataset, evaluating timesteps of 20s, 60s, 80s, 100s, 140s, and 200s to assess their effect on estimation error. The results indicate that shorter timesteps improve accuracy but require larger datasets and increased storage capacity, while longer timesteps yield reasonable errors with a smoother temporal resolution trade-off. These findings offer valuable insights into balancing data granularity and estimation precision, aiding the development of robust SOH prediction methodologies tailored to specific applications.