Rainfall forecasting using artificial neural networks
Department of Mathematics, University of Peradeniya, Peradeniya, Sri Lanka.
By Sathsara Dias, Dr. P. Ekanayake
Employing feed-forward back-propagation architecture based Artificial Neural Networks, a modern approach has been attempted for forecasting the rainfall. This research has been based for making successful predictions from the available data, not on incorporating the physical aspects of the atmosphere nor the actual process of rainfall occurrence. A unique method of short-term forecasting has been attempted with ground level data collected by the meteorological station in Colombo, Sri Lanka.
Three Neural Network models were developed for the ‘one-day-ahead’ model for predicting the rainfall occurrence of the following day for 3 networks. First network was ‘May- September’ which was able to make 95.9% training and predictions with a 77.4% accuracy and overall percentage of 72.9% with accuracy. The layer combination (10 20 20 8) for this was whereas the average correlation coefficient was 0.81999. Next was ‘October- April’ and it made 78.0% training and predictions with a 68.5% accuracy and overall percentage of 57.9% with accuracy. In this there were (10 20 10 8) layer combinations with the average correlation coefficient of 0.81403. The last was ‘No Separation’ which made 97.2% training and predictions with a 26.2% accuracy and overall percentage of 88.8% with accuracy. There were (10 15 15 10 8) layer combinations with the average correlation coefficient of 0.72441 with regards to this. It was determined that by using a dual network for forecasting would yield more accurate results that just running a single network. This was a major learning point of this research.