Data mining techniques applied to statistical prediction of monthly precipitation in Gran Chaco Argentina

Abstract

Data mining techniques are currently a powerful tool to address with the seasonal time-scales forecasting. In this work, neural networks, support vector regression and generalized additive models are considered besides the most commonly used multiple linear regression methodology, to obtain precipitation forecasting models in the area of “Gran Chaco Argentino”. The results indicate that data mining techniques improve forecasts derived from other methodologies, although the efficiency of the different methodologies is highly dependent on the month and the region. The non-linear techniques improve the forecasts and show lower mean square error than the multiple linear regression and support vector regression. The root mean square error is higher east of study area than in the west because precipitation is higher. The coefficient of variation is quite low in all the months in the central and southwest parts of the area. The precipitation interval with the highest probability of occurrence showed a value of 1.5. In addition, the possibility of generating ensemble means of several models and deriving categorical forecasts is a highly advisable alternative for prediction in this region of Argentina. The use of ensemble means is recommended. The derived forecasts improve the dynamic world center models only in some regions of the study area.