Monthly and seasonal drought forecasting using statistical neural networks

Drought forecasting with proper accuracy, notably helps the drought management, and therefore, reduces the damages caused by drought. The aim of this study is to forecast the drought at short-, mid-, and long-term time scales. To this aim, the Standard Precipitation Index (SPI) was calculated on 3, 6, 9, 12, and 24-month time scales based on monthly precipitation data over a 35-year period from 1972 to 2006 in Gorganroud basin. After monitoring the drought, according to the SPI time series and applying six approaches of neural networks, drought forecasting was provided. In the present study, utilized neural networks were Recursive Multi-Step Multi-Layer Perceptron (RMSMLP), Direct Multi-Step Multi-Layer Perceptron (DMSMLP), Recursive Multi-Step Radial Basis Function (RMSRBF), Direct Multi-Step Radial Basis Function (DMSRBF), Recursive Multi-Step Generalized Regression Neural Network (RMSGRNN), and Direct Multi-Step Generalized Regression Neural Network (DMSGRNN). MLP was used in previous studies, in this study this network is applied as a basis for comparing the performance of statistical neural networks (RBF and GRNN). The results based on R2, RMSE, and MAE, showed the forecasting accuracy decreased by increasing lead time of forecasting and increased while SPI time-scale increased. Moreover, recursive models reflected better performance at smaller time scales of SPI, whereas direct models showed better accuracy at longer time scales of SPI. According to the results of drought class forecasting, recursive models performed better than direct models. Generally, the results showed that RBF and GRNN had the best performance in forecasting the drought index and drought class, respectively.