Attention Driven-Chained Transfer Learning for Generalized Sequential State of Charge Forecasting in Vanadium Redox Flow Batteries

Abstract

The increasing integration of renewable energy sources into power grids necessitates efficient energy storage systems to balance supply and demand. Vanadium redox flow batteries (VRFBs) are becoming increasingly popular because of their long lifespan and flexible energy storage capabilities. Central to the effectiveness of VRFBs is the accurate estimation of future state of charge (SOC) levels. However, conventional SOC forecast frameworks suffer from poor generalization capabilities, which restrict their applicability in real-life energy systems. This research introduces a sequential forecast framework that combines multihead self-attention (MHA) with chained transfer learning (CTL) to estimate SOC sequences across multiple temporal horizons. The model performance is evaluated by forecasting SOC levels of the VRFB system operated under various charging and discharging current profiles. The results demonstrate that the change in the VRFB system's operational dynamics significantly reduces the forecast accuracy of conventional frameworks, with the maximum MAE reaching 66%. Compared to the best-performing baseline trained on a linear current profile, the CTL-MHA-gated recurrent unit (GRU) decreased the maximum MAE from 28.7% to below 1.5%. The generalization capability of the proposed framework addresses a critical barrier to the integration of SOC forecast frameworks with smart energy storage systems.