EFFECT OF THE LENGTH OF THE INPUT DATA SERIES ON SIMULATION RESULTS OF AIR TEMPERATURE BY ARTIFICIAL NEURAL NETWORK (ANN) IN NAM BO PLAIN

Abstract

Artificial neural network (ANN) was used to simulate air temperature for Nha Be meteorological station. The data set used includes 5 inputs with an output element of air temperature. ANN structures in Matlab software were designed of 2 hidden layers with 3 levels of neuron number (2, 5 and 8) in each hidden layer and tansig transfer function. Data series lengths varied from 1 month to 48 months giving R from 0.8318 to 0.9673. R values did not follow any certain rule when the length of the data series was no longer than 4 months but bear the same tendency for all three ANN structures when the data was equal or longer than 6 months. With the same ANN structure, the increase in the data series length did not guarantee an increase in R value. The deviation in simulation results from measured data occured more strongly in the sections of the low or high peaks of the data series, especially when the ANN structure has a small number of neurons in the hidden layer.