| description abstract | In this research, the role of climate variability and weather change in short-term streamflows, including extreme event, was investigated in semiarid climates. The deep learning convolutional neural networks (CNN) were modified by incorporating the imperialist competitive algorithm (ICA) and the grey wolf optimizer (GWO) method to improve hourly runoff predictions at multiple scales, ranging from 100 to over 6,000 km2 in the Seybouse Basin, Algeria. The atmospheric reanalysis data set, ECMWF Reanalysis v5 (ERA5) with a 31-km resolution, climate variability indices, and in situ runoff observations were used. The most relevant atmospheric and soil moisture predictors from the reanalysis grids covering the study area were used to represent spatial variability. The prediction performance of the original CNN and modified CNN-ICA and CNN-GWO models were evaluated. The CNN-GWO model outperformed CNN-ICA and the original model in predicting runoff and improved the Nash-Sutcliffe Efficiency score up to 0.99. Results across multiple scales disclose that the models with climate indices as inputs showed higher performance than the models with only atmospheric data as inputs, especially in predicting extreme runoff values in basins with elevations above 670 m, suggesting that climate variability indices need to be considered in flood predictions and infrastructure design in mountainous areas with increasing climate change uncertainties. | |
