Author(s): S. Golian; P. Yavari; H. Ruigar
Keywords: No keywords
Abstract: Drought is one of the natural and climatic disasters and causes abundant damages to human life and natural ecosystems worldwide every year. This study aims to compare the results of meteorological drought forecasting using non-linear auto regressive (NARX) neural network model with those of commonlyused multi-layer Perceptron (MLP) network for Shahrood climate station, located in North-East part of Iran. Different combination of input-output sets with various variables as inputs and SPI with different lag-time as output were tested to determine which combination has the best performance in prediction of future droughts. SPI drought index with 3 and 6 months aggregation period were applied in this study. In general, both MLP and NARX networks had satisfactory performance in prediction of SPI index, but NARX has slightly better performance compared to MLP network for all the models. Also predicting SPI6 resulted in better performance compared to SPI3. Results showed that for both models, using temperature, precipitation, humidity, SPI3, SPI6and SPI12 as input variables and SPI6 with one month lag as output has better performance compared to other input-output combination sets. The optimum calculated performance criteria were CNS (Nash-Sutcliff) =0.954and 0.962, R (correlation coefficient) =0.989 and 0.986, MAE (mean of absolute error) =0.033 and 0.036, RMSE (root mean square error) =0.041 and 0.043 for training and test periods, respectively for NARX network.