Técnicas estadísticas para la modelación y predicción de la temperatura y velocidad de viento en la provincia de Chimborazo.

Abstract

The purpose of the present titration work was to determine the technique that provides better forecasts for modeling meteorological variables, period 2014-2017, for which the variables Temperature and wind speed registered in the meteorological stations of the Center for Alternative Energies and Environment ( CEAA) of the province of Chimborazo, we used the techniques of Box-Jenkins (ARIMA), chaos theory and recurrent neural networks with software support such as R version 3.5.1, Tisean 3.0.1 and the Excel spreadsheet. The imputation of missing data of those bases that did not exceed 20% and considering a minimum adjusted coefficient of determination of R2 = 0.8 was made. Through the techniques: Box-Jenkins it was detected that all the time series had seasonality every 24 lags, identifying SARIMA models and a partial fulfillment of theoretical assumptions, with the theory of chaos models were obtained with better adjustment for the first 48 hours for the temperature, presenting a variation for wind speed due to its instability. In the modeling obtained by recurrent neural networks, predictions were obtained with a greater adjustment to the previous techniques and with less variation in the real versus predicted data. In conclusion, using the U coefficient of theil and the Diebold_Mariano Test, the Box-Jenkins methodology, chaos theory and Recurrent Neural Networks obtained a U of 0.0035, 0.044, 0.020 respectively, considering that the first one is subject to many conditions, the second it is suitable for short-term predictions and the long-term third. 100% of the tests performed identified that recurrent neural networks have greater accuracy and accuracy at 95% reliability.