A Novel State of Charge Estimation Method for Vanadium Redox Flow Battery Based on Cnn-Gru Hybrid Model Under Different Conditions

Abstract

Vanadium redox flow battery (VRFB) is a superior long-duration energy storage technology, offering high safety, ultra-long lifespan, and decoupled scalability of power and capacity. Accurate state of charge (SOC) estimation is critical for enhancing system efficiency, extending service life, and maximize profitability. However, this task remains challenging due to the unique electrolyte flow characteristics of VRFB and dynamic uncertainties during system operation. Previous methods typically rely on prior electrochemical knowledge, and their estimation accuracy deteriorates when flow rates and operating conditions change dynamically. This paper proposes a novel CNN-GRU hybrid model that synergistically integrates the advantage of convolutional neural network (CNN) in extracting complex features with the capability of gated recurrent unit (GRU) in modeling temporal relationships. The CNN is first employed to extract features from test data that include voltage, current, and flow rate. Subsequently, GRU is used to process temporal relationships and establish nonlinear mappings with SOC, thereby enabling accurate and stable SOC estimation. This method accounts for the impact of dynamic flow rate variations and eliminates the need to build a battery model. Training and validation of the model use test data collected under varying flow rates and operating conditions. Comparisons with other methods demonstrate significant superiority, achieving an average RMSE of 2.79%, MAE of 2.28%, MAXE of 7.5%, and R² of 98.52%. Additionally, results show that once trained, the model can achieve SOC prediction at approximately 47μs per sample with high computational efficiency, making it particularly suitable for battery management systems (BMS).

Keywords: Vanadium redox flow battery, State-of-Charge, Hybrid data-driven model, Flow rate, Convolutional neural network, Gate recurrent unit