<![CDATA[Integrating photovoltaic (PV) power plants into smart grids requires reliable short-term forecasting, particularly under highly variable tropical atmospheric conditions. Conventional univariate temporal models, such as LSTM, struggle to anticipate abrupt cloud cover changes as they rely solely on historical electrical data. This study proposes a hybrid multimodal CNN-LSTM framework integrating 16-channel Himawari-9 spatio-temporal satellite imagery and real-time inverter electrical parameters via cross-climate transfer learning and Solar Polar Image Normalization (SPIN). Multimodal alignment was achieved using linear temporal and bilinear spatial interpolations across 9,622 synchronized 5-minute samples from Universitas Pamulang, while a dedicated multimodal adapter enabled partial weight transfer from the Stanford SKIPP’D dataset. Experimental results demonstrate superior performance, achieving an MAE of 3,73 kW, RMSE of 5,61 kW, a Forecast Skill of +5,98%, and mitigating extreme Q95 error to 12,37 kW compared to persistence and univariate LSTM baselines. With a CPU inference latency of 15,58 ms per sample, this architecture provides high accuracy and computational efficiency for real-time tropical grid operations.]]>
Copyrights © 2026
