Ensuring a generalizable machine learning model for forecasting reservoir inflow in Kurdistan region of Iraq and Australia

Correct inflow prediction is a critical non-engineering measure for ensuring flood control and increasing water supply efficiency. In addition, accurate inflow prediction can offer reservoir planning and management guidance since inflow is the major input into reservoirs. This study aims at generalizing a machine learning model for forecasting reservoir inflow. Daily, weekly, and monthly inflow and rainfall time-series data have been collected as two hydrological parameters to forecast reservoir inflow using a machine learning method, namely, support vector regression (SVR). Four different SVR kernels have been applied in this study. The kernels are radial basis function (RBF), linear, normalized polynomial, and sigmoid. Two scenarios for input selection have been implemented. Dokan dam in Kurdistan region of Iraq and Warragamba Dam in Australia were selected as the case studies for this research. For the purpose of generalization, the proposed models have been applied to two countries with a different climate condition. The findings showed that daily timescale outperformed weekly and monthly, while RBF outperformed the other SVR kernels with root-mean-square error (RMSE) = 145.7 and coefficient of determination (R2) = 0.85 for forecasting daily inflow at Dokan dam. However, RBF kernel could not perform well for forecasting daily inflow in Warragamba dam. The results showed that the proposed machine learning model performed well at Kurdistan region of Iraq only, while the result for Australia was not accurate. Therefore, the proposed models could not be generalized.