Effects of input variables preprocessing in SPI prediction using artificial neural network and wavelet transformation

Drought is a natural event that can bring considerable damage to human life. The prediction of drought can play an important role in the water resources management during the drought periods. Currently, ANNs have shown great ability in forecasting non-linear time series. On the other hand, wavelet transform improves the resolution using decompositions of an original time series to sub-signals. In this study, three hybrid models including perceptron neural network wavelet (MLP-W), recurrent network wavelet (TR-W) and time lag recurrent network wavelet (TLRN-W) were presented for drought prediction and then, the Standardized Precipitation Index was predicted using these models for 12 months ahead in Yazd meteorological station. In addition, in order to evaluate the effect of wavelet transforms on performance of hybrid models, the results of hybrid models were compared with the results of the single ANN models using statistical criterion including R, RMSE, and MAE. Finally, the results of hybrid models showed a higher correlation coefficient and lower error in comparison with single ANN models.The correlation coefficient, RMSE and MSE in the best hybrid model were calculated to be 0.977, 0.05, and 0.020, while these values in the best single ANN model (TLRN) were. 895 and 0.07 and 0.020, respectively. In general, it was found that wavelet transforms could improve the performance of neural networks in drought prediction.