Machine learning assisted prediction for hydrogen production of advanced photovoltaic technologies

Abstract

The photovoltaic (PV) water electrolysis method currently stands as the most promising approach for green hydrogen production. The rapid iteration of photovoltaic technologies has significantly affected on the technical and economic evaluation for photovoltaic hydrogen production. In this work, the photovoltaic hydrogen production of three most advanced silicon photovoltaic technologies is systematically compared for the first time under the climatic conditions of the Kucha region. All-weather stable hydrogen production control system with optimal charging and discharging strategies is constructed to realize stable and efficient hydrogen energy production. Seven machine learning (ML) algorithms are used to forecast the performance in power generation and hydrogen production of a 100 ​MW photovoltaic hydrogen production and energy storage (PH-S) system throughout its operational life. The long short-term memory (LSTM) algorithm exhibits the best performance, achieving mean absolute error (MAE) of 0.0415, root mean square error (RMSE) of 0.0891, and coefficient of determination (R2) of 0.8402. In terms of cost-effectiveness, heterojunction with intrinsic thin layer (HJT) PV technology achieves the lowest levelized cost of electricity (LCOE) and hydrogen (LCOH) at 0.025 $/kWh and 6.95 $/kg, respectively. According to the sensitivity analysis, when the cost of proton exchange membrane electrolysis (PEMEC) reduced 50%, the LCOH for PH-S system decreased 21.40%. This study provides valuable insights for the practical implementation of large-scale photovoltaic hydrogen production and cost reduction in PH-S systems.