Multi-station runoff-sediment modeling using seasonal LSTM models

Abstract

In this study two types of seasonal long-short-term memory (LSTM) artificial neural networks, named sequenced-LSTM (SLSTM) and wavelet-LSTM (WLSTM), were presented to model runoff-sediment process of three gauging stations, located in Missouri and Upper Mississippi regions in both daily and monthly scales. For this purpose, twenty-year observed streamflow and suspended sediment load (SSL) data were employed in both daily and monthly scales. The proposed seasonal models have full profits of classic LSTM model in time series processing and handle sole LSTM model’s weaknesses in failing to capture seasonal information of the process which usually exist in hydro-climate time series. The proposed models enhanced the long-short autoregressive dependency of runoff-sediment data by taking into consideration of very long seasonal dependency of data. The obtained outputs indicate the outperformance of proposed seasonal LSTM models to the classic LSTM and feed forward neural network models in test step up to about 25% and 28% in daily and monthly scales, respectively.