Long-Lead Rainfall Forecasting, Using Dynamic Neural Networks: Case Study of Western Part of Iran

Application of Temporal Neural Networks in long-lead forecasting of seasonal rainfall over western part of Iran is presented. Three approaches including application of time delay operators, recurrent connections as well as a hybrid method are used to design Artificial Neural Network (ANN)-based models. Climatic variables including the difference between Sea Level Pressure (SLP) at different characteristic locations in Middle-East and Europe are used as the predictors of the models as well as the persistence between rainfall time series. The characteristic locations include some parts of Mediterranean and Black sea, Greenland, and Azores. The models are calibrated based on 31-year data of historical regional rainfall and are validated by a cross validation procedure. Further, an Auto Regressive Moving Average with eXogenous input (ARMAX) are used as the baselines for assessing the performance of the dynamic networks. The results demonstrated that all temporal neural networks especially time delay recurrent neural network perform significantly better than statistical ARMAX models in long-lead seasonal rainfall forecasting in west of Iran.

Keywords:

Rainfall , Forecasting , Temporal Neural Networks , ARMAX