Less is more: Robust prediction for fueling processes on hydrogen refueling stations

Abstract

The monitoring of hydrogen refueling stations (HRSs) ensures the safety of their operations as well as optimal fueling performance. For a H70-T40 dispenser, a fueling process is required to control the temperature to be below 85 °C; the pressure to be under 70 MPa; and the final state-of-charge (SOC) to be between 95% and 100%. Table-based or MC (total heat capacity) formula-based fueling protocols are traditionally used to achieve such control. In this paper, we propose using a machine learning model to predict the key parameters of fueling processes: the final SOC, the final temperature, and the final pressure in the vehicle tank. To handle outliers and noise in real operation, we adopt a two-stage method. In the first stage, after clustering fueling processes using soft dynamic time warping, a small number of fueling processes are selected from a large amount of historical data. In the second stage, based on initial and current operating conditions, the final SOC, temperature, and pressure of fueling processes are predicted using three models: least absolute shrinkage and selection operator (LASSO), Gaussian process regression (GPR), and robust regression. The experiments on real operational data collected from four hydrogen refueling stations show that the robust regression model achieves better performance than LASSO and GPR for three out of the four stations, and that the robust regression model captures the normal states of regular operation. The computational time of the robust regression model is also scalable for real-time operation. Our study provides a feasible machine learning model for predicting the key fueling parameters, which facilitates the optimization of HRS operation.