EFFECTS OF INPUT PARAMETER SELECTION ON ARTIFICIAL NEURAL NETWORK SIMULATION RESULTS FOR AIR EMPERATURE AT CAN THO AND NHA BE METEOROLOGICAL STATIONS

Abstract

This study assessed the effects of different input data sets including several basic meteorological factors on air temperature simulation results using Artificial Neural Network (ANN) at Can Tho and Nha Be meteorological stations. ANN structures with different number of neurons provided overall correlation coefficient value (R) in a quite wide range (0.5156 ÷ 0.9658). The results also showed that ANN structure with a single hidden layer of 6 neurons using tansig transfer function is suitable to simulate air temperature at the above-mentioned stations. Simulation accuracy was higher with the input data containing the typical air temperature measured at the same day at another station and the Day of year, thereby revealing strong macroscopic climate influence to air temperature at both stations. The importance of input parameter selection was demonstrated via a smaller variation in overall R value caused by switching ANN structures compared to that from changing parameters of the input data set.