<![CDATA[The variability of photovoltaic (PV) power under tropical weather conditions complicates one-hour-ahead forecasting and operational energy planning. This study evaluates deep learning-based PV power forecasting by integrating historical PV output from a university in Malang, Indonesia with NASA POWER, Solcast, BMKG, and time features. Five data scenarios were assessed using Persistence, LSTM, and GRU, while hybrid architectures and source- and architecture-level ablation tests were evaluated on the complete multi-source configuration. Timestamp continuity was explicitly audited, and 178 holdout sequences that crossed temporal gaps were removed. GRU with the complete configuration achieved the lowest baseline RMSE of 0.218839 kW, with MAE of 0.110686 kW, nRMSE of 6.230669%, and R² of 0.955550. Relative to Persistence, RMSE and MAE decreased by 49.60% and 54.36%, respectively. LSTM with the Solcast scenario produced a nearly identical RMSE of 0.218992 kW, indicating that the benefit of additional data sources was architecture-dependent and non-monotonic. CNN-LSTM was the best hybrid model but remained 3.26% worse in RMSE than GRU, while the tested attention mechanism increased RMSE by 10.02%. Four-fold limited walk-forward validation supported GRU with the lowest mean RMSE of 0.232213 ± 0.019500 kW. The final forecasts were also translated into conceptual Energy Management System decision-support recommendations. Because same-day BMKG daily summaries were used in the main configuration, the findings are primarily interpreted as a retrospective evaluation.]]>
Copyrights © 2026
