Region‐Resolved Machine‐Learning Forecasting of Photovoltaic‐Driven Hydrogen Production Potential in Uganda: A Long Short‐Term Memory–Prophet Comparison

Abstract

This study develops a region‐resolved machine‐learning framework for forecasting photovoltaic‐derived hydrogen‐production potential in Entebbe, Kampala, Mbarara, and Gulu, Uganda. Hourly meteorological data from 21 December 2023 to 20 December 2024 were used for model development, while the subsequent 45 days were reserved for independent chronological testing. Global horizontal irradiance, air temperature, wind speed, and relative humidity were selected through correlation analysis, principal‐component assessment, and chronological validation. Photovoltaic electrical‐energy availability was estimated first and then converted into hydrogen‐production potential using a screening‐level constant‐efficiency electrolyzer model. Long short‐term memory and Prophet forecasts were compared using R 2 , MAE, RMSE, and nRMSE. Long short‐term memory (LSTM) consistently outperformed Prophet, achieving R 2 values of 0.931–0.958 compared with 0.829–0.949 for Prophet. Relative to Prophet, LSTM reduced RMSE by 15.6%–37.2% and MAE by 43.2%–50.0%. Kampala and Gulu exhibited the highest monthly hydrogen‐production potential, approximately 41 000 kg H 2 km −2 of active PV area month −1 , corresponding to about 2.16 GWh km −2 month −1 of PV electricity. The results demonstrate that regional meteorological variability affects both production potential and forecastability, while LSTM provides the more reliable basis for comparative site screening, reserve assessment, and preliminary scheduling of PV–hydrogen systems.