In recent decades, researchers have used various methods, including artificial neural networks, to predict floods. These networks have several subsets. Long Short-Term Memory (LSTM) is an improved version of Recurrent Neural Networks (RNN). Learning long-term dependencies for traditional recurrent neural networks is difficult. The LSTM model, by introducing a memory cell that can retain information for a long time, overcomes this problem. This network has been introduced as an efficient method for flood prediction. In this study, the average flow values of the Baghestan and Tezerjan hydrometric stations in the Fakhrabad area of Yazd province during the period 1398-1370 have been used. The precipitation values of the top ten stations in the study period showed a mild decreasing trend. To evaluate the performance of the LSTM model in predicting flow, the coefficient of determination (R2) and RMSE between the observed flow values and the predicted flow by the LSTM model at the Baghestan and Tezerjan stations on a daily and three-month time scale was calculated. The results showed that among the models used, the LSTM model used for predicting the daily stramflow at the Tezerjan station showed the highest performance with R2 and RMSE values of 0.78 and 0.43, respectively. Overall, the results showed that due to the irregular distribution of precipitation in dry areas, LSTM models have relatively acceptable performance in predicting floods in dry areas.