استفاده از روش های یادگیری عمیق و الگوریتم حافظه کوتاه بلند مدت (LSTM) در مدلسازی اطلاعات حوزه های آبخیز

One of the most important behaviors of a watershed is the relationship between the amount of received precipitation, the produced runoff, and the sediment associated with this runoff. In this research, an attempt was made to retrieve information using deep learning techniques and LSTM algorithm for three selected hydrometric stations in Iran, including Qazaqli, Sira, and Rodak. Two scenarios were designed for modeling. In the first scenario, to predict the next day's sediment (at time t), the time series of the previous day's sediment (at time t-1) was inputted at each step. In the second scenario, to predict the next day's sediment (at time t), the time series of sediment in the past 30 steps (from time 30-t to t) and the time series of discharge at the same time mentioned in the 30 steps before were used. Despite the fact that LSTM model is an improved version of RNN, in the second scenario, RNN performed better with its own limitations for all three stations with the following indices for Qazaqli station: MSE= 6.54E+07 and RMSE=8092.51 (milligrams per liter), for Sira station: MSE = 9.38E+07 and RMSE=9688.62 (milligrams per liter), and for Rodak station: MSE= 1.75E+07 و RMSE=13232.44 (milligrams per liter). Additionally, among the three hydrometric stations, the LSTM algorithm in the Sira station showed lower errors in both scenarios compared to the other stations with the indices of MSE = 7.3E+08 and RMSE=27044.37 (milligrams per liter) in the first scenario and MSE= 8.7E+08 and RMSE=29517.5 in the second scenario.