Wind Power Generation Forecast by Coupling Numerical Weather Prediction Model and Gradient Boosting Machines in Yahyalı Wind Power Plant

Published: 02-12-2020| Version 5 | DOI: 10.17632/wg9pkrp3rv.5
Cem Özen,
Umur Dinç


A low-resolution WRF model has run which has exact same spatial resolution with GFS of NCEP which is initial and boundary conditions data and extracted data from the four grid points of this model which surround the related wind farm are used as predictor variables to the gradient boosting machines. Thus; the downscaling process is basically done by the coupling of the low-resolution numerical weather prediction model and machine learning algorithm. The results of this hybrid model are also compared with the results of a WRF model which has a higher spatial resolution in terms of both statistical measures and required time. As a result of this; the study claims a novel approach for not only increasing the accuracy of the wind power generation forecasts but also reducing the computational expense. Since wind energy increases its share in installed energy capacity thanks to its maturity in terms of technology and decreasing costs; wind power generation forecasts with high performance becomes very crucial to have not only the security of the electricity supply of the countries all around the world but also prevent energy imbalance penalties for the wind farm owners. Besides, this study shows that a better performance than the downscaling processes of the numerical weather prediction models do can be achieved in shorter periods with machine learning algorithms. Thus; this study also highlights that the coupling of machine learning algorithms with the numerical weather prediction models can play an effective role in the future in terms of atmospheric sciences. Consequently; this work provides basically not only increasing the accuracy of the wind power generation forecasts when it is compared with the numerical weather prediction models but also decreasing the computational expense by doing the downscaling process of numerical weather prediction models with gradient boosting machines.


Steps to reproduce

Data can be used to reproduce the results. While raw low and high-resolution wrf model outputs for the last test day as samples; wps and input namelists of them are shared. Besides, turbine based production data which provide us to calculate total wind farm production and available number of turbines are shared and KML file which shows license area and the coordinates of the 22 wind turbines of the Yahyalı Wind Farm can also be found at this directory. On the other hand, test and train data which are ready to use as input to the machine learning algorithm are also shared and can be found under the "exported" subdirectory under the LowResolutionWRF directory. Moreover, results of high resolution WRF model can also be found under the "exported" subdirectory of HighResolutionWRF directory. Furthermore; final GBM model of the grid search is shared and it is found at the "r_h2o_gbm_model" directory. While R programming language has been used as post processing works, building a gradient boosting machine model and data visualization of the data; python has been used for exporting the atmospheric variable from the wrf model output.