Performance prediction of wind energy-driven green hydrogen production: A case study based on optimized machine learning models for abandoned mine sites

In addressing the critical need for decarbonizing energy systems and utilizing underutilized land, abandoned mining sites offer a distinctive opportunity to develop integrated wind-to-hydrogen facilities capable of supplying both renewable electricity and green hydrogen. In this study, two machine-learning models, an artificial neural network (ANN) and a long short-term memory (LSTM) network, whose hyperparameters are selected using particle swarm optimization (PSO), are used to concurrently predict wind turbine power and hydrogen production at three island mining locations in South Korea. Both models underwent fine-tuning using PSO hyperparameter searches and were evaluated using the RMSE, MAE, MSE, and R 2 metrics across the training (2014–2021), validation (2022–2023), and blind test (2024) phases, achieving R 2 > 0.9998 in all scenarios. A techno-economic assessment determined the levelized cost of hydrogen (LCOH), calculated as the net present value of all system costs divided by the discounted hydrogen output over 30 years, to be $2.584/kg at Damuldo-Ri, $2.272/kg at Gageodo-Ri, and $2.845/kg at Socheong-Ri. Hydrogen storage requirements for an 800 bar composite tank ranged from 4.7 to 5.2 metric tonnes peak (124–137 m 3 ) under monthly drawdown, decreasing to 1.1 to 1.2 tonnes (28–31 m 3 ) for weekly cycling. Fuel cell performance, based on a PEM stack with latent-heat recovery, resulted in annual average electrical power outputs of 96.79 kW (stack), 13.79 kW (recovered), and 110.58 kW (total) at Damuldo-Ri (851.75, 121.38, 973.13 MWh), 111.73 kW, 15.92 kW, and 127.65 kW at Gageodo-Ri (981.11, 139.82, 1120.93 MWh), and 87.27 kW, 12.44 kW, and 99.70 kW at Socheong-Ri (767.12, 109.30, 876.42 MWh). CO 2 emission reductions through gasoline substitution and wind-derived electricity ranged from 277.24 to 354.59 tCO 2 avoided annually, with wind energy contributing 701.14–896.74 tCO 2 and fuel-cell backup adding 52.16–57.51 tCO 2 to the savings. These findings demonstrate that PSO-tuned neural models can accurately predict combined wind and hydrogen outputs with less than one percent error, whereas integrated economic and environmental analyses confirm that repurposed mining sites can provide cost-effective green hydrogen and significant CO 2 reduction. • Models predict wind power and H 2 at three mining sites (<1 % annual error). • PSO-LSTM cuts wind power RMSE by ∼25 % at Gageodo-Ri and Socheong-Ri vs ANN. • Levelized hydrogen cost falls to $2.27–$2.85 per kg over 30-year project life. • Wind energy yields 2–3 × more CO 2 savings than hydrogen substitution across sites.