ABSTRACT Accurate short‐term solar power forecasting is essential for grid stability under high photovoltaic (PV) penetration, particularly in tropical monsoon climates where irradiance is highly variable. This paper proposes a dual‐branch hybrid deep learning framework that integrates Himawari‐9 satellite imagery with ground‐based meteorological and PV data. A CNN branch extracts spatial cloud features from image sequences, while an LSTM branch captures temporal dependencies from numerical time‐series; their outputs are fused to generate multi‐horizon nowcasting predictions. The framework is validated on an 11‐month dataset from Hanoi, Vietnam, and benchmarked against four baselines: a CNN‐LSTM trained only on numerical data, a standalone CNN, a standalone LSTM applied to image sequences, and a naive persistence model. Results demonstrate that the hybrid model consistently outperforms all baselines across both 30‐ and 60‐min horizons. At 30 min, it reduces MAE by approximately 49% and RMSE by approximately 41% relative to the best deep‐learning baseline (numerical‐only CNN‐LSTM), and achieves 30% lower MAE than the naive persistence reference, while maintaining stable accuracy under volatile cloud conditions (reducing MAE by 65% compared to the numerical‐only model) where competing models fail. Distributional analyses (ECDF, parity plots, violin‐box) confirm minimized variance, reduced outliers and the absence of systematic underestimation at peak generation hours. These findings highlight the operational relevance of hybrid spatio‐temporal deep learning for solar integration in challenging climates. Beyond Vietnam, the approach offers a scalable pathway to enhance reserve allocation, storage scheduling and grid reliability in regions pursuing aggressive renewable energy targets.
Nguyen‐Duc et al. (Thu,) studied this question.