Development of a new wavelet-based hybrid model to forecast multi-scalar SPEI drought index (case study: Zanjan city, Iran)

Drought forecasting plays a vital role in managing drought and reducing its effects on agricultural systems and water resources. In the present study, three machine learning models including Gaussian Process Regression (GPR), Cascade Neural Network (Cascade-NN), and Multilayer Perceptron (MLP) neural network and their combination with the discrete wavelet transform were used to forecast Multi-scalar Standardized Precipitation Evapotranspiration Index (SPEI) (SPEI3, SPEI12, and SPEI24) 1 to 6 months ahead. It was done in Synoptic Station of Zanjan in Iran. Those meteorological data that was collected during 57 years (1961–2017) was used. The data related to the early 38 years (67%) was considered as train data, and the data related to the last 19 years (33%) was considered as test data. The results that have been obtained from this study showed that models based on wavelet have caused a high improvement in model performance in case of anticipating multi-scalar SPEI. Comparing different mother wavelets (db4, db8, sym8, coif5, and dmey) proved the dmey wavelet’s superiority. Also, a comparison of wavelet-GPR, wavelet-MLP, and wavelet-Cascade-NN models showed that in most cases, the GPR-based model could provide better results in forecasting. By increasing the forecasting interval from 1 to 6 months ahead, the accuracy of the model decreased. In the SPEI3 index, the R2 (determination coefficient) value decreased from 0.992 in the 1-month ahead forecast to 0.797 in the 6 months ahead forecast. In the SPEI12 index, the R2 value decreased from 0.996 in the 1 month ahead forecast to 0.940 in 6 months ahead forecast, and in the SPEI24 index, R2 values decreased from 0.993 in the 1 month ahead forecast to 0.962 in 6 months ahead forecast.