Wind gust forecasting by post-processing the WRF model outputs using ANN

Strong and highly variable winds and gusts are major hazards to infrastructure, properties, and life. Consequently, accurate prediction and timely detection of wind gust intensity have always been a focus of interest for earth scientists and weather forecasters.

In this study, The WRF (Weather Research and Forecasting) post-process diagnostic of wind gusts (WPD method) was utilized to predict non-convective wind gust speeds using the direct outputs of the WRF model. To improve the prediction accuracy of this method, the results were post-processed using an artificial neural network (ANN). Multiple different ANN algorithms were examined to achieve the most accurate predictions possible. The results were evaluated using observational data extracted from 32 synoptic stations across Iran during the time period from 2014 to 2018.

The results indicate that employing a multilayer perceptron ANN with a hybrid structure, consisting of one input layer comprising five parameters (10 m wind speed, sea level pressure, temperature, relative humidity, and predicted wind gust speed obtained from the WPD method), one hidden layer with a sigmoid activation function and 12 neurons, one output layer with a linear activation function and using the BR (Bayesian Regularization) training algorithm, significantly improve the accuracy of the WPD wind gust speed prediction method. The RMSE for wind gust speed prediction has decreased from 3.68 m/s (WPD method) to 1.88 m/s for the validation dataset. Additionally, there were considerable improvements of 50 %, 74 %, and 17 % in the MAE, MSE, and R², respectively.