Feature selection and prediction of Gusty wind with multilayer perceptron neural network (MLP) at the airport auto station

In this paper, in the first step, the present weather of METAR reports of the year 2013 in Mehrabad synoptic station was studied and the period with most occurrences of the instability producing the Gusty wind was identified. This period is from January to June of every year. Then, all data of selected period, except the data of Gusty wind direction and speed, were normalized to interval 0.1–0.9. The considered data for training, testing and validation were 60%, 20% and 20%, respectively. The related features of Gusty wind direction and speed were selected from 58 features recorded by 3 sensors located on the runway. The Mehrabad runway direction is from the east to the west with 4000 meters long and 45 meters wide. The sensor No. 29 was on the east end of band, the sensor No. 11 was on the west edge of the band, and location of the mid sensor was on the middle of band which its distance from the band is 600 meters to the north direction. The feature selection methods in this study are mutual information (MI) with the Maximum-Relevance Minimum-Redundancy criterion (filter type) and Sequential Floating Forward Selection (SFFS) (wrapper type) with the k Nearest Neighbors (kNN) algorithm. Selected features for Gusty wind speed at each band are the maximum and mean wind speed in 2 and 10 minutes, and the momentary wind speed by the MI method. The selected feature by SFFS method is the wind direction deviation in past 10 minutes on band No. 11 and mid band, momentary pressure on mid band and maximum wind speed in 10 minutes on band No. 29. For Gusty wind direction by first method, the selected features are minimum, mean and maximum wind direction in 2 minutes, minimum and mean wind direction in 10 minutes and momentary wind direction on band No. 29. Selected features with second method are the wind direction deviations in past 10 minutes on the band No. 29 and mid band, and the mean sea level pressure and mean wind direction in 10 minutes on band No. 29. In the final step, these selected features were used as inputs of the multilayer perceptron neural network in different modes such as: layer number, neuron number, learning rate and threshold value for weight of neuron. The model output results were compared to predict the Gusty wind direction and speed and the best model was selected. The results show that to predict the wind speed, the best model is a multilayer perceptron neural network with four layers: input layer with 4 neurons, two hidden layers with 4 neurons in the first layer and 2 neurons in the second layer and 1 neuron in the output layer; learning rate of 0.1 and initial weight neurons of 0.5. For predicting the wind direction, the best model has four layers, 6 neurons in the first and second layers and 3 neurons at the third layer and one neuron at the fourth layer with the same learning rate and initial threshold. The MLP performance is better in predicting the Gusty wind speed.