Long short-term memory and metaheuristic algorithm- based deep learning model approach for forecasting green hydrogen production

This study proposes a hybrid forecasting framework that integrates Long Short-Term Memory (LSTM) networks, Kernel Principal Component Analysis (KPCA), and Particle Swarm Optimization (PSO) to predict green hydrogen production from High Concentration Photovoltaic (HCPV) systems coupled with PEM electrolyzers. The model is applied to hourly time-series data from several cities in southern Morocco. Extensive data preprocessing, feature extraction and selection, dimensionality reduction, and PSO-based hyperparameter tuning (hidden units, dropout rate, learning rate, and batch size) were performed to enhance hourly predictive performance. Findings show strong annual electricity and hydrogen outputs across the six sites, with Boujdour achieving the highest values (35.49 GWh and 736.7 t), followed by Smara and Laayoune. The LSTM-KPCA-PSO model outperforms the baseline LSTM, achieving lower RMSE and higher R², particularly in Boujdour, where training RMSE decreased from 2.068 to 1.275 and testing RMSE from 7.173 to 6.278. Likewise, R² increased from 0.9998 to 0.9999 during training and from 0.9972 to 0.9978 during testing.