Linear and Nonlinear Statistical Downscaling for Rainfall Forecasting over Southeastern Brazil

Abstract

In this work linear and nonlinear downscaling are developed to establish empirical relationships between the synoptic-scale circulation and observed rainfall over southeastern Brazil. The methodology uses outputs from the regional Eta Model; prognostic equations for local forecasting were developed using an artificial neural network (ANN) and multiple linear regression (MLR). The final objective is the application of such prognostic equations to Eta Model output to generate rainfall forecasts. In the first experiment the predictors were obtained from the Eta Model and the predictand was rainfall data from meteorological stations in southeastern Brazil. In the second experiment the observed rainfall on the day prior to the forecast was included as a predictor. The threat score (TS) and bias, used to quantify the performance of the forecasts, showed that the ANN was superior to MLR in most seasons. When compared with Eta Model forecasts, it was observed that the ANN has a tendency to forecast moderate and high rainfall with greater accuracy during the austral summer. Also, when the observed rainfall of the previous day is included as a predictor, the TS showed the best performance in continuous rain and well-organized meteorological systems. On the other hand, in the austral winter period, characterized by slight rain, the ANN showed better forecasting ability than did the Eta Model. The obtained results also suggest that in the austral winter rainfall is more predictable because convection is less frequent, and when this occurs the forcing is dynamic instead of thermodynamic.