Multi-Source Data Integration for Photovoltaic Power Forecasting in Tropical Indonesia

Abstract

Photovoltaic (PV) power forecasting in tropical environments is challenged by rapid cloud-driven irradiance variability and changing meteorological conditions. This study systematically evaluates one-hour-ahead PV power forecasting using historical PV data from the Universitas Muhammadiyah Malang PV system integrated with NASA POWER, Solcast, BMKG, and temporal features. The objective is to determine when multi-source information improves forecasting and whether added model complexity yields consistent gains. Five input scenarios were evaluated with Persistence, LSTM, and GRU, complemented by hybrid-model comparisons, source and attention ablation, four-fold limited walk-forward validation, and BMKG availability sensitivity. GRU-S5 achieved the lowest holdout RMSE of 0.2188 kW, 49.60% lower than Persistence; however, LSTM-S3 reached 0.2190 kW, demonstrating that more sources do not guarantee better performance across architectures. Source ablation showed that removing NASA POWER or BMKG increased GRU RMSE by 4.52% and 4.27%, whereas the Solcast contribution was marginal and seed- and metric-sensitive. CNN-LSTM was the best hybrid but remained 3.26% worse in RMSE than GRU-S5, and the tested attention mechanism increased RMSE by 10.02%. Walk-forward validation yielded the lowest mean RMSE for GRU at 0.2322 ± 0.0195 kW, although model rankings varied across folds. Overall, effective multi-source forecasting depends on source relevance, temporal availability, and architecture rather than feature count alone. The forecasts further support conceptual EMS decision support, while same-day BMKG use makes the main evaluation primarily retrospective.