Accurate state of charge prediction for real-world battery systems using a novel dual-dropout-based neural network

Accurate prediction of the state of charge is critical to the safety and durability of battery systems in electric vehicles. This paper proposes a novel multi-step SOC prediction method for real-world battery systems using the gated recurrent unit recurrent neural networks, which fully considers the influences of the environment and driving behaviors on the prediction performance. A novel dual-dropout method is proposed to prevent overfitting and optimize training efficiency. The first dropout is based on Pearson correlation analysis approach. It extracts five actual vehicle parameters that are strong and implicitly correlated with predictive SOC as model inputs, including recorded SOC, pack voltage, vehicle speed, temperature of probe, and brake pedal stroke value. A random dropout function is constructed as the second dropout to decrease the network density and improve efficiency, which is applied to the state information passing process of the model. Furthermore, the training samples are constructed by deriving the yearlong operation data of an electric taxi. The optimal model framework and hyperparameters are discussed and determined. Verified by six sets of randomly selected vehicular operation data, the results show that the proposed method can perform real-time 5-min SOC prediction with maximum error of 0.86%. • A novel multi-step SOC prediction method based on GRU-RNN is proposed. • A dual-dropout overfitting prevention method is explored by binning and random dropout. • A real-world dataset is derived from an electric taxi as the training and testing data. • Accurate multi-step real-time prediction of battery state of charge is obtained. • Stability, robustness, and superiority are verified using real-world operation data.