Solar forecasting preprint reports lower errors with combined neural models

Researchers report in an arXiv preprint that a hybrid model outperformed statistical, machine learning and deep learning baselines in solar power forecasting tests. The system produces forecasts for 30-minute, 2-hour, 6-hour, 12-hour and 24-hour horizons within one pipeline. It combines a Temporal Fusion Transformer, which learns short-term changes, with N-BEATS, which captures longer-term trends and seasonal patterns. An attention mechanism adjusts how the two models contribute to predictions. The researchers report mean absolute percentage errors below 5% across the tested datasets and deployment on a cloud platform for real-time forecasting. Evaluation used multi-year data from the Desert Knowledge Australia Solar Centre.