چکيده

Introduction: Increasing water demand and water pollution due to the development of agricultural, urban and industrial activities have caused environmental problems all over the world. The significant increase in water pollution and the diversity of various urban, agricultural and industrial pollutants made the qualitative management of water resources inevitable. Short-term and long-term accurate forecasts of river quality parameters are essential for designing hydraulic structures, irrigation planning, optimal utilization of reservoirs and environmental planning. Given the stochastic characteristics of the hydrological events, forecasting the future status of surface waters is always associated with uncertainties. The purpose of the present study was to investigate the performance of two types of artificial neural networks, namely MLP and GMDH, combined with discrete wavelet transform (DWT), to forecast two important quality parameters, electrical conductivity (EC) and sodium adsorption ratio (SAR) at Zayandeh-Rood River in 1, 2 and 3 months ahead.   Material and methods: In this study, water quality data (EC and SAR) of Zayandeh-Rood River at Zaman Khan Station was used from 1363 to 1384. From 21 years of data, 15 years (approximately 70%) were used for training and 7 years (30%) were used to test the developed models. Two types of mother wavelet dmey and db4 were evaluated. Statistical parameters such as RMSE and R² were used to evaluate the performance of the models.   Results and discussion: The results showed that the use of discrete wavelet transform improves the performance of the models. Various combinations of input data (various delays) and two types of mother wavelets were evaluated. The results showed that wavelet-MLP and wavelet-GMDH hybrid models outperform single MLP and single GMDH models at all forecasting intervals. The results of the single MLP and GMDH models were only effective in forecasting SAR one month ahead but practically could not forecast two and three months later. In the EC parameter, the MLP and GMDH models performed better. Also, the results showed that the use of annual time lags does not increase the accuracy and in some cases even reduces it. The study of the types of mother wavelets also showed that the dmey wavelet is the most suitable wavelet type to forecast EC and SAR qualitative parameters. The comparison between wavelet-MLP and wavelet-GMDH models showed the relative superiority of the former model. By increasing the forecast period from one month to three months ahead, the accuracy of the models decreased. This decrease in precision was higher in forecasting SAR parameter, e.g. in the one month forecast, R² was 0.936 and in the 3 months ahead forecasts it was reduced to 0.516. In the EC parameter, the R² fell to 0.641 in 3 months ahead forecasting.   Conclusion: The results of this study can be used as a basis for future planning for water quality. It is suggested that the model presented in this study should be considered in other rivers. Also, the combination of other artificial intelligent models such as ANFIS and SVM with wavelet transform can be evaluated.