Solar irradiance forecasting has become increasingly important as photovoltaic capacity grows globally, yet predicting hourly solar power remains challenging due to the complex interplay of atmospheric conditions, seasonal patterns, and regional climate variations. A new study presents a comprehensive machine learning framework that combines multiple neural network architectures to deliver robust forecasting performance across diverse locations.
The research employs a three-stage approach. First, gradient boosting algorithms identify the most relevant meteorological and radiation variables for each specific site, improving model interpretability and reducing computational overhead. Second, the core prediction engine chains three complementary architectures: convolutional neural networks extract local temporal patterns, bidirectional long short-term memory networks capture dependencies in both temporal directions, and transformer layers apply attention mechanisms to refine predictions.
Testing across seven NASA satellite-derived datasets and one utility-scale ground station demonstrated strong performance, with the best site achieving 94.7% explained variance and mean absolute error below 0.13 kilowatt-hours per square meter. Comparisons with conventional methods including seasonal ARIMA, Prophet, and simpler neural networks confirmed competitive advantage. Notably, the model maintained performance consistency across different climatic regions, addressing a critical gap in prior single-site studies.
SHAP analysis revealed that predictions relied on physically meaningful variables—cloud cover, aerosol optical depth, and seasonal indicators—validating the model's decision-making logic. Ablation studies confirmed each architectural component contributed meaningfully; removing any stage degraded accuracy.
For grid operators and utilities, this framework offers practical value. Improved solar forecasting enables better unit commitment decisions, reduces reserve margin requirements, and supports efficient integration of distributed photovoltaic systems. The reproducible, open-source methodology allows adaptation to site-specific conditions without extensive retraining. As variable renewable penetration accelerates, such forecasting tools become essential infrastructure for maintaining grid stability and optimizing operational costs across increasingly solar-dependent power systems.
