(211e) Data-Driven Predictive Model and Optimization Based on Machine Learning on Steam Reforming Process

Hydrogen has recently been in the spotlight as its potential as a future eco-friendly mobility energy source. Although the on-site steam methane reforming (SMR) process is widely used for hydrogen refueling stations, the difference between the thermal efficiencies of commercial products and the theoretical maximum is very large. To maximize thermal efficiency, numerous studies regarding mathematical modeling have been conducted. However, mathematical modeling has a disadvantage in that it is difficult to build high-fidelity models of complex equipment and processes. Besides, the on-site SMR process has complicated reactors and processes. Therefore, in this study, a predictive model based on machine learning is developed to overcome the difficulty. Firstly, a data pre-processing step is performed to improve the data quality, removing outliers and noise of 485,710 actual operation datasets. Next, the optimal number of hidden layers, number of hidden nodes, optimizer, and activation function are determined to develop a high accuracy deep neural network model (R² value of 0.9987). The predictive model estimates six variables including steam temperature, reformate gas flow rate, and its CO, CO₂, CH₄, and H₂ compositions. To determine the optimal operating conditions, a total of 387,420,489 cases are calculated, considering both nine operating variables and five constraints. In conclusion, the thermal efficiency is maximized to 85.6%, higher than those of commercial products.