Improved streamflow prediction accuracy in Boreal climate watershed using a LSTM model: A comparative study

Streamflow plays a vital role in water resource management and environmental impact assessment. This study is a novel application of the Long Short-Term Memory (LSTM) model, a type of recurrent neural network, for real-time streamflow prediction in the Upper Humber River Watershed in western Newfoundland. It also compares the performance of the LSTM model with the physically based SWAT model. The LSTM model was optimized by tuning hyperparameters and adjusting the window size to balance capturing historical data and ensuring prediction stability. Using single input variables such as daily average temperature or precipitation, the LSTM achieved a high Nash-Sutcliffe Efficiency (NSE) of 0.95. In comparison, the results show that the LSTM model delivers a more competitive performance, achieving an NSE of 0.95 versus SWAT’s 0.77, and a percent bias (PBIAS) of 0.62 compared to SWAT’s 8.26. Unlike SWAT, the LSTM model does not overestimate high flows and excels in predicting low flows. Additionally, the LSTM successfully predicted daily streamflow using real-time data. Despite challenges in interpretability and generalizability, the LSTM model demonstrated strong performance, particularly during extreme events, making it a valuable tool for streamflow prediction in cold climates where accurate forecasts are crucial for effective water management. This study highlights the potential of the LSTM model’s application to water resource management.Author Summary: Accurate streamflow prediction is essential for managing water resources, particularly in cold Boreal climates like Newfoundland, Canada. This study explores the use of an LSTM machine learning model to predict streamflow in the Upper Humber River Watershed. We compared it to the widely used SWAT model and found that the LSTM model performed better, accurately predicting both high and low flow events. With fewer input features, the LSTM model effectively captures seasonal variations, extreme weather, and real-time forecasts. This study shows the value of combining advanced ML techniques with traditional models to improve streamflow predictions, especially in the context of increasing climate variability.