Forecasting Seasonal and Annual Rainfall Based ...

Rainfall plays a key role in hydrological application and agriculture in wet climatic regions. Lack of short-run rainfall Forecasting is considered as a significant impediment for scheduling the root zone moisture preparation. Although many mathematical techniques are available for use, basic concerns remain unsolved such as simplicity, high accuracy, real time use in many stations of a region, and the low availability of inputs. In this study, a nonlinear modeling with Gamma Test (GT) has been presented to solve some of the mentioned problems. Forecasting seasonal and annual rainfall with the variables of four years lagged rainfall data and geographical longitude, latitude and elevation has been performed in the North of Iran during 1956-2005. The results show that how well the GT works in conjunction with the nonlinear modeling techniques for rainfall forecasting. The applied nonlinear modeling techniques are Local Linear Regression (LRR), Dynamic Local Linear Regression (LRR), and three separate Artificial Neural Networks (ANN) using Back Propagation Two Layer, Broyden-Fletcher-Goldfan-Shanno (BFGS), and the Conjugate Gradient training Algorithms. The training and testing data are partitioned by random selection from the original data set. The Gamma Test not only yields the best input combination, but also how good the model’s performance is relative to the best achievable result so that it is possible for a modeler to decide whether any further effort is worthwhile to improve the existing model. The study results demonstrates that developed models based Local Linear Regression (LRR) technique have better performance comparing with ANN models. Also, developed ANN model based on Back Propagation Two Layer training Algorithm is preferred because of its better performance compared with the other ANN models.