Comparative Study of Different Discrete Wavelet Based Neural Network Models for long term Drought Forecasting

Recently, coupled Wavelet transform and Neural Networks models (WANN) were extensively used in hydrological drought forecasting, which is an important task in drought risk management. Wavelet transforms make forecasting model more accurate, by extracting information from several levels of resolution. The selection of an adequate mother wavelet and optimum decomposition level play an important role for successful implementation of wavelet neural network based hydrologic forecasting models.The main objective of this research is to look into the effects of various discrete wavelet families and the level of decomposition on the performance of WANN drought forecasting models that are developed for forecast drought in the Algerois catchment for long lead time. The Standard Precipitation Index (SPI) is used as a drought measuring parameter at three-, six- and twelve-month scales. Suggested WANN models are tested using 39 discrete mother wavelets derived from five families including Haar, Daubechies, Symlets, Coiflets and the discrete approximation of Meyer. Drought is forecasted by the best model for various lead times varying from 1-month lead time to the maximum forecast lead time. The obtained results were evaluated using three performance criteria (NSE, RMSE and MAE).The results show that WANN models with discrete approximation of Meyer have the best forecast performance. The maximum forecast lead times are 36-month for SPI-12, 18-month for SPI-6 and 7- month for the SPI-3. Drought forecasting for long lead times have significant values in drought risk and water resources management.

Automated summary

Recently, the coupled Wavelet transform and Neural Networks models (WANN) have been widely used in hydrological drought forecasting, which is crucial for drought risk management. The study focuses on the effects of different discrete wavelet families and the level of decomposition on the performance of WANN models in drought forecasting for long lead times. The results show that WANN models with discrete approximation of Meyer perform the best, with maximum forecast lead times of 36 months for SPI-12, 18 months for SPI-6, and 7 months for SPI-3. Drought forecasting for long lead times is highly significant for drought risk and water resources management.