Thesis (M.Sc.)--Chulalongkorn University, 2009

Rainfall prediction generally requires reliable hydrological models as well as relevant information such as meteorological and geographical data. In this thesis, the prediction is, first, investigated by a conventional artificial neural network (ANN) model on accounts of the preceeding events of rainfall data and climatological variables. Since precipitation behavior can be categorized into 2 time-periods; wet and dry period, ANN models for each time period are, then, implemented in order to improve network accuracy. A model based on artificial neural network (ANNs) and wavelet decomposition is another proposed learning tool predicting consecutive daily rainfalls on accounts of the preceding events of rainfall data only. In the combining model of ANN and wavelet decomposition, two sets of wavelet coefficients, for which one pattern represents detail information of rainfall data and the other acts as a smoothing filter, are extracted for the ANNs. A back-propagation neural network is used in the learning and knowledge extraction processes for all models. The methodologies are tested on rainfall data from five stations in the eastern coast of southern Thailand; ThaSae district in Chumphon province, Kanchanadit district in Surat Thani province, Muang district in Nakhon SriThammarat province, Muang district in Phatthalung province and Hatyai district in Songkhla province. From the past historical records of Thai Meteorological Department and Royal Irrigation Department, these study area are vulnerable to flood disaster. The proposed model based on artificial neural network (ANNs) and wavelet decomposition is capable of forecasting daily rainfall up to 4 days in advance with accuracy of R-square = 0.8887 and RMSE = 4.2306 mm.