A neural network-based rainfall prediction with different input selection methods

Planning and managing water resources in arid and semi-arid countries such as Iran highly depends on the long-term prediction of precipitation. In this study, weather signals and artificial neural networks have been used for predicting long-term precipitation. In order to do this task, climatic data and meteorological data with 3 to 12 months lead-times have been used as inputs to predict precipitation for 3, 6, 9 and 12 months periods in 6 selected stations across Iran. Stations have been selected in such a way that covers different types of weathers and climates in Iran. These stations consist of Abadan, Babolsar, Birjand, Kermanshah, Yazd and Zahedan. A genetic algorithm (GA) and self-organized neural network (SOM) along with the application of winGamma software have been used comparatively as input selection methods to choose the appropriate input variables . Based on the results, out of 96 predictions performed at all stations, in 31 cases, GA method, in 31 cases, winGamma method, and in 34 cases SOM method have the best results in comparison with other two methods. According to this, as a general assumption, it can be concluded that at least for the selected stations in this work, the SOM method is more reliable than the other two methods, and can be used to make predictions for future applications as a reliable input selection method.Moreover, among different climatic signals, Pacific Decadal Oscillation (PDO), Trans-Niño Index (TNI) and Eastern Tropical Pacific SST (NINO3) are the most repetitive indices for the most accurate forecast of each station

Keywords:

Precipitation Prediction , Large Scale Climatic Signals , Self-Organized Artificial Neural Networks , Input Selection Method , Genetic Algorithm