Accurate solar irradiance forecasting is essential for photovoltaic (PV) scheduling, grid stability, and long-term energy planning. Nevertheless, most of the previous research has focused on short-term horizons, relied on heterogeneous ground-based inputs, and lacked systematic feature validation and decision-oriented model evaluation. This study develops a robust hybrid deep-learning framework integrated with Multi-Criteria Decision Making (MCDM) to enhance both forecasting accuracy and interpretability. Using a 20-year (2004–2023) NASA satellite-based dataset covering four distinct U.S. climates, six hybrid architectures (LSTM–CNN–MLP, CNN–LSTM–GRU), and three single-stream baselines were benchmarked following Chi-square feature selection. The proposed MCDM framework, TOPSIS, assesses the models in three decision scenarios, i.e. the accuracy-dominant, efficiency-dominant, and balanced trade-off with an emphasis on the RMSE, MAE, R2, computational time, parameters, and CPU energy. The findings indicate that the hybrid models are always more effective than individual networks, and LSTM–CNN–MLP had R2 = 0.98 and MAPE = 2% at all the horizons. The framework shows 2−10 times smaller RMSE and greater cross-climate strength, which offers a transferable and low-cost answer to satellite-only, multi-horizon solar forecasting. The results place the proposed model as a rational decision and scalable instrument of operational PV management and renewable energy management in the world.
Chang et al. (Mon,) studied this question.