Computationally sustainable hybrid deep learning for month-ahead solar–wind power forecasting in two Saudi cities

This paper implements artificial intelligence for month-ahead renewable power forecasting by fusing synchronized solar and wind information from the Sakaka Solar Plant and Dumat Al-Jandal Wind Farm in Saudi Arabia. Ten hybrid deep-learning architectures are evaluated under a common temporal framework using mean absolute error, root mean squared error, mean absolute percentage error, coefficient of determination, training time, and computational carbon dioxide emissions. The convolutional neural network-long short-term memory model with attention achieves the highest predictive accuracy, with a mean absolute error of 80, a root mean squared error of 129, a mean absolute percentage error of 6.6%, and a coefficient of determination of 0.87. The Temporal Fusion Transformer provides the strongest accuracy-sustainability compromise under the 70:30 Sustainable Artificial Intelligence score, while the wavelet-convolutional neural network-gated recurrent unit model records the lowest training footprint at 768 grams of carbon dioxide equivalent. A 10,000-draw perturbation sensitivity analysis of the error and emission values confirms that the preferred deployment model is decision-dependent rather than universally dominant: the Temporal Fusion Transformer is most frequently top-ranked under the central weighting, whereas carbon-heavier or carbon-lighter alternatives become preferred as decision weights change. The contribution is therefore system-level: the framework integrates cross-resource data fusion, leakage-aware long-horizon evaluation, carbon-aware model selection, Pareto analysis, and explicit decision-robustness testing. Reported carbon reductions refer to model computation and should not be interpreted as measured reductions in grid curtailment or citywide operational emissions.