Machine learning-based energy optimization for on-site SMR hydrogen production. (15th September 2021)

• Record Type: Journal Article
• Authors: Lee, Jaewon; Hong, Seokyoung; Cho, Hyungtae; Lyu, Byeonggil; Kim, Myungjun; Kim, Junghwan; Moon, Il
• Is Part Of: Energy conversion and management. Volume 244(2021)
• Publication Date: 2021-09-15
• Subjects: On-site hydrogen production -- Steam methane reforming -- Data preprocessing -- Neural network -- Machine learning -- Energy optimization; Direct energy conversion -- Periodicals; Energy storage -- Periodicals; Energy transfer -- Periodicals; Énergie -- Conversion directe -- Périodiques; Direct energy conversion; Periodicals; 621.3105
• DOI: 10.1016/j.enconman.2021.114438
• Languages: English
• ISSNs: 0196-8904

Abstract

Highlights: DNN data-driven model for steam methane reforming is developed. Data preprocessing techniques were applied to improve the quality of datasets. Four hyperparameters were optimized to obtain an accurate prediction model. Process optimization was conducted with 387, 320, 489 cases under five constraints. The operating conditions were optimized to achieve a thermal efficiency of 85.6%. Abstract: The production and application of hydrogen, an environmentally friendly energy source, have been attracting increasing interest of late. Although steam methane reforming (SMR) method is used to produce hydrogen, it is difficult to build a high-fidelity model because the existing equation-oriented theoretical model cannot be used to clearly understand the heat-transfer phenomenon of a complicated reforming reactor. Herein, we developed an artificial neural network (ANN)-based data-driven model using 485, 710 actual operation datasets for optimizing the SMR process. Data preprocessing, including outlier removal and noise filtering, was performed to improve the data quality. A model with high accuracy (average R 2 = 0.9987) was developed, which can predict six variables, through hyperparameter tuning of a neural network model, as follows: syngas flow rate; CO, CO2, CH4, and H2 compositions; and steam temperature. During optimization, the search spaces for nine operating variables, namely the natural gas flow rate for the feed and fuel, hydrogen flow rate for desulfurization, water flow rate and temperature, air flow rate, SMR inlet temperature and pressure, and low-temperature shift (LTS) inlet temperature, were defined and applied to the developed model for predicting the thermal efficiencies for 387, 420, 489 cases. Subsequently, five constraints were established to consider the feasibility of the process, and the decision variables with the highest process thermal efficiency were determined. The process operating conditions showed a thermal efficiency of 85.6%.