Hydrological time series play an important role in engineering problems such as flood and drought management and the safe design of water structures. Although hydrological time series are generally considered to be random variables that repeat in certain periods (e.g. annually), the series has an internal dependence. It is very difficult to represent the internal dependence of such time series deterministically due to the complex nature of the events. For this reason, machine learning techniques, which have been a popular black box method in recent years, can be used effectively in modelling hydrological time series. Such time series models can be used to fill in missing data in hydrological time series, such as river flows, or to predict future data when sufficient data is available. In this study, we considered using the flow series of a particular river, observed over a long period of time, in order to predict these time series using different machine learning techniques. Two different Long Short-Time Memory (LSTM) approaches (open-loop and closed-loop forecasting) were used as deep learning techniques. The method monitored the performance of three different optimizer. The results were compared and an attempt was made to decide on the best method for estimating the streamflow series.