Rainfall forecasting from multiple point sources using neural networks

Weather forecasting has been one of the most challenging problems around the world for more than half a century. Not only because of its practical value in meteorology, but it is also a typically "unbiased" time-series forecasting problem in scientific researches. The paper describes the methodology to short-term rainfall forecasting using neural networks. It extends a previous study relying on observational data from a single point station to multiple point sources with time-series weather records in the Hong Kong region. Preprocessing procedures were important for this neural network modeling which was based on a backpropagation architecture. This involved variable transformation, classification and the use of genetic algorithms for input selection. Compared with previous studies on a single point source using a similar network and others like radial basis function networks, learning vector quantization and naive Bayesian network, the results are very promising. This neural-based rainfall forecasting system is useful and parallel to traditional forecasts from the Hong Kong Observatory.