Forecasting Droughts using Artificial Neural Networks

Abstract The use of artificial neural networks as a tool to forecast droughts in Sri Lanka is presented. Predictions were made using the Standardized Precipitation Index (SPI) as the drought monitoring index. Monthly rainfall recorded at 13 climatological stations covering both the wet and dry zones over a long time period have been used as the input to train and test the neural networks. The analysis covers rainfall data recorded over the time span 1870 to 1980. The SPI was computed by fitting a probability density function to the frequency distribution of the monthly precipitation records of each station. The developed neural network model was tested for SPIs with time windows of 1-6 months. For SPI-3 computed with a 3 month time window, the average correlation coefficient was found to be 0.90 with the lowest being 0.84 for Nuwara Eliya and highest being 0.94 for Batticaloa and Jaffna. In general, the accuracy of the predictions was higher for the stations in the dry zone compared to the stations in the wet zone. The model predictions were superior for the period from May to July (which is the first part of the South-West Monsoon season) compared to the rest of the year. The accuracy of the predictions increased with the length of the window used in computing the SPI values. The results of this work shows that neural network models trained on SPI can be used to forecast water scarcity.