Simulation of Qanat Discharge in Balade Ferdows Using Machine Learning Algorithms Using Drought Indices

Abstract

In recent decades, the country has faced decreased rainfall, increased temperatures, and a severe decline in groundwater resources. Qanats, as one of the most important indigenous water supply systems in arid and semi-arid regions, have been severely affected by climate change and the recent droughts. In this regard, employing advanced machine learning models can play a key role in developing reliable forecasting systems to support climate change adaptation planning. The objective of this research is to simulate the monthly discharge of the Balade Qanat complex in Ferdows County using a set of machine learning models, including single algorithms such as XG Boost, SVR, Random Forest, and Gradient Boosting, as well as an advanced ensemble approach, Stacking. This simulation uses climatic, hydrological data, and drought indices over 10 years. The dominant approach in modelling is comparing the performance of individual models against the final ensemble model. The obtained results showed that under the region’s variable climatic conditions, the Stacking ensemble approach exhibited a significantly stronger performance than single models like XG Boost. The Stacking model was selected as the optimal model, achieving the highest coefficient of determination (R²) and the highest Kling-Gupta Efficiency (KGE = 0.93, R² = 0.92) with the lowest Root Mean Square Error (RMSE = 12.21). This superior performance emphasises the capability of Stacking models in reducing variance and correcting systematic biases of individual models when dealing with the complex and nonlinear behaviour of Qanat discharges. It is concluded that the Stacking model, due to its ability to extract complex nonlinear patterns and improve generalisation, is a superior management tool for decision-making in the sustainable exploitation of Qanat water resources in water-stressed climates.