Comparison of data-driven techniques for daily streamflow forecasting

Abstract

Four artificial intelligence methods are compared for streamflow forecasting. The models are tested using 20 years of daily streamflow values in seven basins of the Zagros Mountain Range, Iran. The models considered in the study are artificial neural networks (ANNs), Artificial Neural Networks trained with Ant Colony Optimization for continuous domains (ACO_ℝ–ANN), Adaptive Neuro-Fuzzy Inference System (ANFIS) and Multigene Genetic Programming (MGGP). The performances of the models are measured by the root mean square error, the coefficient of determination (R²) and the Nash–Sutcliffe model efficiency. Depending on the basin, ANN, ANFIS or MGGP is the best performing method. None of the methods outperforms the others for all the basins. Overall, the best-performing model is ANN and the worst is ACO_ℝ–ANN. The physical and climate characteristics of the basins influence the models’ performances.