Wavelet-transform based artificial neural network for daily rainfall prediction in southern Thailand

Rainfall prediction generally requires reliable hydrological models as well as relevant information of meteorological and geographical data. In this paper, a model based on artificial neural networks (ANNs) and wavelet decomposition is proposed as a learning tool to predict consecutive daily rainfalls on accounts of the preceding events of rainfall data. 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. The methodology is tested on rainfall data from five stations in the south of Thailand: tha Sae district in Chumphon province, Kanchanadit district in Surat Thani province, Muang district in Nakhon Si Thammarat 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 areas are vulnerable to heavy rainfall distribution and flood disaster. The proposed network is capable of forecasting daily rainfall up to 4 days in advance with accuracy of R² = 0.8819 and RMSE = 4.6912 mm.