Stochastic Rainfall Forecasting for High Tropical Andean Regions

Abstract:

High Tropical Andean systems play a fundamental role in rainfall water generation for the region. Therefore, management and decision making require more accurate quantification and forecasting. However, the latter is inherently difficult, due to the non-stationarity nature of climate series, and the differences in climate conditions. Nonetheless, stochastic models, such as the Autoregressive Integrated Moving Average (ARIMA, or an extension thereof with Seasonal terms SARIMA), and Dynamic Harmonic Regression (DHR), are very frequently used techniques for forecasting purposes. Interesting Differencing structures and Random Walk (RW) processes are embedded in ARIMA and DHR for handling non-stationarity. However, since in every particular application several RW processes and structures are suitable for DHR and ARIMA respectively, they have to be carefully chosen from all of them using appropriate statistical criteria (Identification). An applied study carried out over a South-Ecuadorian region reveals the following results: 1) According to Akaike Information Criteria (AIC), Autoregressive Moving Average (ARMA, which are simpler ARIMA models) perform best. 2) The best performing RW process embedded in DHR techniques is the simple well-known Autoregressive of order one (AR [1]). 3) After Identification process, DHR and ARMA models perform virtually equal according to the Coefficient of Determination (R2), Root Mean Square Error (RMSE); except for BIAS in which slight differences are shown, although both models underestimate and overestimate peaks and lows respectively.

Año de publicación:

2017

Keywords:

Fuente:

google