Multi-steps drought forecasting in arid and humid climate environments: Development of integrative machine learning model

In the current study, the Standardized Precipitation Evaporation Index (SPEI) was forecasted using a combination of the empirical wavelet transform (EWT), discrete wavelet transforms (DWT), extended Kalman filter (EKF), two models of multilayer perceptrons (MLP), and group method of data handling (GMDH) neural networks. Two synoptic stations of Tabriz (semi-arid climate) and Rasht (humid climate) covering data period (1987–2019) were selected for forecasting. 70% of the data was used for model training and 30% for validation. Three forecasting horizons (1, 3, and 6 months ahead SPEI) were investigated. Autocorrelation function and partial autocorrelation function were used to determine the optimal inputs to the models. The outcomes of the present study showed that in both stations, the combination of machine learning models with two types of wavelet transforms (EWT and DWT) compared to the standalone models improved the performance of the forecasting (correlation coefficient, R = 0.9980, root mean square error, RMSE = 0.0483 for Tabriz Station and R = 0.9988, RMSE = 0.0521 for Rasht Station). A comparison of the EWT and DWT wavelets showed that the EWT had better performance in all forecasting intervals. By raising the forecasting interval from one month to six months, EWT performance was more evident than DWT performance. In 6-month forecasting horizon, the DWT had almost no effect on model performance improvement. In both stations, the combination of the EWT and MLP-EKF model had the best performance in forecasting SPEI drought index.