Akademik Veri Yönetim Sistemi
International Conference on New Technologies, Development and Application, 2022, Sarajevo, Bosna-Hersek, 23 - 25 Haziran 2022, cilt.472 LNNS, ss.813-822
© 2022, The Author(s), under exclusive license to Springer Nature Switzerland AG.Prediction of streamflow is important in managing surface water resources, especially in agricultural production. The streamflow process is a function of different variables, primarily the spatial and temporal distribution of meteorological parameters, catchment and river physical characteristics. The complex nonlinear relationship between streamflow and influential variables can be expressed through various methods such as stochastic, probabilistic, empiric or black box models. In this research, the daily streamflow of the Nalli River, located in Ankara, Turkey, has been modelled by using four different architectures of the recurrent neural network (RNN). LSTM, BiLSTM, GRU and Simple RNN architectures were selected. The input dataset includes daily streamflow, precipitation and their antecedent values. The dataset is divided into two subsets: training (70%) and testing (30%) datasets of a total of 6939 daily measurements (1997–2015). All four RNN variants were coded in the Python programming language. The results of the methods were compared with each other based on the correlation coefficient (CC), Nash–Sutcliffe model efficiency coefficient (NS), mean absolute error and mean square error. The results showed that the LSTM method is more accurate than other deep learning methods, as it can estimate the input current in the training period with a precision (CC) of 97% and in the test period with a precision of 96%. According to the results, it can be stated that due to their accuracy, deep learning methods can be used to predict the flow of the Nalli River.