Streamflow modelling and forecasting for Canadian watersheds using LSTM networks with attention mechanism

Abstract This study investigates the capability of sequence-to-sequence machine learning (ML) architectures in an effort to develop streamflow forecasting tools for Canadian watersheds. Such are useful inform local and region-specific water management flood related activities. Two powerful deep-learning variants Recurrent Neural Network were investigated, namely standard attention-based encoder-decoder long short-term memory (LSTM) models. Both models forced with past hydro-meteorological states daily meteorological data a look-back time window several days. These tested 10 different watersheds from Ottawa River watershed, located within Great Lakes Saint-Lawrence region Canada, economic powerhouse country. The results training testing phases suggest that both able simulate overall hydrograph patterns well when compared observational records. Between two models, attention model significantly outperforms all watersheds, suggesting importance usefulness mechanism ML architectures, not explored hydrological applications. mean performance accuracy on unseen data, assessed terms Nash–Sutcliffe Efficiency Kling-Gupta is, respectively, found be 0.985 0.954 these Streamflow forecasts lead times up 5 days demonstrate skillful above benchmark 70%. encoder–decoder LSTM, mechanism, is modelling choice developing systems