Forecasting streamflow in Asi River using long-short term memory networks

River flow estimation is necessary in order to control water resources and to ensure sustainable use. Long-term flow forecasting is of great importance for water resources planning and management in both the short and long term. Traditional instantaneous flow measurements in river basins take a lot of time and require high costs. Instead of instantaneous measurements, artificial intelligence programs that can predict flow with today's technology reveal very important developments. Long-short term memory (LSTM) neural networks from deep learning neural networks, which support developments and are widespread in use, are adopted to predict the flow of river flows on daily and monthly scales. In this study, the LSTM model was created with daily flow data from Afrin flow measurement station (FMS) no D19A006, which is within the boundaries of the Asi river basin, and the predictive performance of the model was analyzed. The flow values used in the model between 2006-2018 were used as test data set. For the study, was tested with MSE, Huber and LogCosh loss functions using Adam and Adamax optimizer. When determining the best forecast model, statistical evaluation criterion were determined by comparing the observed data and forecast models. Considering the evaluation criterion, the LogCosh loss function of the Adamax optimizer was found to have better performance and R² 0,9949 was achieved. Also, according to the model analysis, LSTM analysis of daily flow forecasting with different scenarios has shown that the performance impact can be expanded and help ensure the sustainability of water resources.