Wind Speed Downscaling of the WRF Model at Subkilometer Scale in Complex Terrain for Wind Power Applications

• Published in: IEEE journal of selected topics in applied earth observations and remote sensing 2024, Vol.17, p.9139-9177
• Main Authors: Paola, Francesco Di, Cimini, Domenico, De Natale, Maria Pia, Gallucci, Donatello, Gentile, Sabrina, Geraldi, Edoardo, Larosa, Salvatore, Nilo, Saverio Teodosio, Ricciardelli, Elisabetta, Ripepi, Ermann, Viggiano, Mariassunta, Romano, Filomena
• Format: Article
• Language: English
• Subjects: Algorithms; Artificial neural network (ANN); Artificial neural networks; Atmospheric modeling; Computation; Computer simulation; Computing costs; Cost analysis; Datasets; downscaling; Electric power generation; Forecasting; Large eddy simulation; Neural networks; Numerical models; Numerical weather forecasting; Orography; Prediction models; Predictive models; remote sensing; resolution enhancement; Sea surface; Spatial discrimination; Spatial resolution; Weather; Weather forecasting; weather research and forecasting (WRF); weather research and forecasting large eddy simulation (WRF-LES); Wind fields; Wind forecasting; Wind power; Wind power generation; Wind speed
• ISSN: 1939-1404; 2151-1535
• DOI: 10.1109/JSTARS.2024.3386629

Mesoscale numerical weather prediction models are frequently utilized for wind speed analysis and forecasting in the planning and support of wind power generation. However, high computational costs only allow for routine use up to a kilometer scale, which is sometimes too coarse to support onshore wind power generation in areas with complex orography. To address this, an algorithm was developed in southern Italy to downscale the wind fields output using the weather research and forecasting (WRF) model for the first 250 m above ground level. The algorithm is based on artificial neural networks (ANNs) and uses the WRF model outputs on a 1.2 km regular grid, and the land surface height and orientation on a 240 m regular grid to downscale wind fields to a 240 m regular grid. To train the ANNs, a WRF simulation dataset in large eddy simulation (LES) mode was developed. Particular attention was paid to defining the ANN architectures and analyzing inputs to mitigate overfitting risk while maintaining manageable computation costs. The evaluation of outcomes conducted using independent test datasets from WRF-LES simulations reveals that the wind speed root-mean-square difference (RMSD) is 0.5 m/s over land and 0.2 m/s over the sea surface, respectively, at a spatial resolution of approximately 800 m. These figures are lower than the RMSD values of 1.6 m/s over land and 1.0 m/s over the sea surface, accompanied by a spatial resolution of 1.8 km, which were obtained through comparison with the spline interpolation method.