Effects of stacking LSTM with different patterns and input schemes on streamflow and water quality simulation

Abstract

Abstract Streamflow and water quality parameters (WQs) are commonly forecasted by mechanism models and statistics models. However, these models are challenged due to computational complexity, redundant parameters, etc. Therefore, a stacking Long short-term memory networks (LSTM) model with two patterns and different input schemes was applied to simulate streamflow and eight WQs in this study. The results showed that sliding windows was detected as the more stable pattern for both forecasts. The accuracy of predicting streamflow using only meteorological inputs was limited especially with low-volume flow. Whereas, the prediction of WQs with three input variables (i.e., meteorological factors, streamflow, other influential WQs) was reliable reaching an average relative error (RE) below 17%. When adding historical data into the input dataset, both accuracies could be increased close to benchmarks of the Delft 3D model. Our study documents that the LSTM model is an effective method for streamflow and water quality forecasts.