Abstract:

Direct Normal Irradiance (DNI) is a key variable for solar resource assessment and the design of concentrating solar power plants, yet it is less frequently measured than Global Hor-izontal Irradiance (GHI) due to high costs and operational constraints. To address this gap, a novel two-stage deep learning framework is proposed that first forecasts GHI and then reconstructs DNI through an advanced separation model, eliminating the reliance on dedicated DNI measurements. The framework leverages bidirectional recurrent neural networks (BiLSTM/BiGRU) to capture temporal dependencies, combined with a wrapper-based Mutual Information (WMI) feature selection method to optimally integrate diverse radiometric and atmospheric quantities. Validated on 15-minute data representing arid, dust-prone, tropical, and temperate climates, the WMI-DL consistently and significantly outperformed four conventional empirical separation methods (Erbs, DISC, DIRINT, Engerer4) across RMSD, MAD, MBD, and R² metrics. At Gobabeb and Desert Rock arid sites, RMSD is reduced to 10.4 W/m² and 20.7 W/m², respectively, with R² of 96.6% and 90.5%, while at Tamanrasset, under arid and frequently dusty conditions, RMSD reaches 18.5 W/m² with an R² of 93.7%, markedly outperforming empirical models. Even in cloudy temperate climates, the model achieves RMSD values of 30.5 W/m² in Bermuda and 36.3 W/m² in Palaiseau, with R² ≈ 91%, demonstrating robustness across diverse atmospheric conditions. While direct DNI forecasting with BiGRU achieves slightly higher accuracy, the WMI-DL framework provides a cost-effective, adaptable, and robust solution for high-fidelity DNI estimation, outperforming both conventional separation methods and ECMWF reanalysis benchmarks in regions lacking direct measurements.

Publisher's Version