Evaluation of two methods of forecasting wind gust speed in Iran and post-processing of results using artificial neural network

Atmospheric currents, known as winds, are among the most important fields of study in different disciplines of science. One of the most important characteristics of wind is gustiness. Wind gust, among many other characteristics of the wind field, is studied extensively due to severe impacts that it may have on many aspects of human socio-economic activities. There are several models to predict wind gust speed. The results of these models always contain random and systematic errors that reduce the accuracy of predictions due to the lack of topographic resolution as well as the deficiencies of different physical schemes in the models. Consequently, post-processing is the most important process in the course of simulation and prediction using different types of models. Artificial neural network is one of the available tools that may be used to reduce errors of models by matching their outputs and observations.     The aim of this study was to evaluate the performance of two models and artificial neural network in forecasting wind gust in Iran. First, a study was designed to examine two methods of the non-convective wind gusts forecasting, i.e., the UK Meteorological Office (MOA) and WRF post-process diagnostic of wind gusts (WPD) performances. To investigate the performace of two methods, 1880 cases of non-convective wind gust observations of 32 synoptic stations in Iran, between 2013 and 2018, were studied. Four RMSE, MAE, MSE and R were used to measure the performace of those two methods. The results for WPD and MOA were 3.89, 3.07, 15.2, 0.66 and 4.37, 3.43, 19.1, 0.55, respectively. The results showed that the WPD method performed better than the MOA method. To post-process the wind gust forecasts with an artificial neural network, a feedforward multilayer perceptron with the back-propagation learning algorithm was designed. The model had a hybrid structure with a sigmoid activation function for the hidden layer and a linear transfer function in the output layer. Three training algorithms were used in the implementation of the model. Various combinations of normalized output variables of the WRF were used as input for network training and the target was observational wind gust speed. Seventy percent of the data were used for training, fifteen percent for testing and fifteen percent for validation.     The results showed that the best way to combine the input parameters is to use 10m wind, sea level pressure, temperature and relative humidity resulting from the output of the WRF model and the wind gust speed resulting from both methods mentioned above. Also, the best algorithm for neural network training was the Levenberg-Marquardt algorithm. Finally, the implemented artificial neural network was able to improve the results of both wind gust speed prediction methods (WPD and MOA). Due to the relatively higher accuracy of the WPD method compared with MOA method in predicting the wind gust speed in Iran, the artificial neural network that assumed the prediction of this method as input, was more accurate than MOA method (RMSE, MAE, MSE and R were 2.05, 1.6, 4.21, 0.83 and 2.37, 1.86, 5.2, 0.77, respectively).