CL-Kansformer model for SOC prediction of hydrogen refueling process in fuel cell vehicles

The rapid compression of hydrogen during refueling causes the temperature inside the hydrogen storage tanks (HSTs) to rise sharply. Hydrogen refueling stops when the temperature exceeds the boundary limit of 85 °C. Overcharging or undercharging of the state of charge (SOC) at the end of hydrogenation can easily occur. Real-time control of mass flow rate through SOC prediction feedback during refueling is an effective solution. Therefore, accurate prediction of SOC during refueling is crucial. For this purpose the article collects and screens 70 MPa real hydrogen refueling data and builds the CL-Kansformer prediction model. This model uses convolutional Long Short-Term Memory (LSTM) as the input layer to extract the multidimensional features of the data and encode them globally. The Kolmogorov-Arnold Networks (KAN) is used to replace the Multi-Layer Perceptron (MLP) in Transformer to improve the prediction accuracy. The prediction results show that the average absolute error of this model is less than 0.2, the root-mean-square error is less than 0.18, and the coefficient of determination is more than 0.99. It can predict the SOC of the hydrogen refueling process more accurately than a single algorithmic prediction model. This model provides a research basis for constructing an intelligently controlled rapid hydrogen refueling system. •Proposed deep learning model to predict SOC of hydrogen refueling processes.•Collected and filtered real hydrogen refueling data as model inputs.•Use multiple algorithm fusion approach to build SOC prediction models.•The proposed predictive model outperforms single-algorithm predictive models.