ASSESSMENT OF CLIMATE DATA AUTO - CORRELATION AT TWO METEOROLOGICAL STATIONS IN NAM BO PLAIN USING PARTIAL AUTOCORRELATION FUNCTION (PACF) AND ARTIFICIAL NEURAL NETWORK (ANN)

Abstract

This study used Artificial Neural Networks (ANN) and Partial autocorrelation function (PACF) graphs to evaluate auto - correlation level in time series of climate factors measured during 2014 - 2017 period at Can Tho and Nha Be meteorological stations. The ANN structure has 2 hidden layers of the form n - 25 - 25 - 1, Levenberg-Marquardt backpropagation training algorithm with "tansig" transfer function for hidden layers. The results had demonstrated that auto - correlation problem exists at a statistically significant level for all climate factors considered, but is often damped (with a maximum interval of 5 - 7 days, especially up to 30 days for total rain). Except for total rainfall factor, the maximum correlation coefficient value (R_max) of 1000 independent ANN simulations was relatively effective to evaluate the degree of correlation for the first lagged (or earlier) day data, although it was still not able to clearly determine the influence of further days. For ANN detection to be effective, the PACF value of a day lagged (or earlier) to the date considered will probably have to be higher than 0.4 and at least 2 times higher than the next largest PACF value.