LSTM for Streamflow Forecasting: From Simulation to Real-Time Flood Warnings

​Can an LSTM actually forecast streamflow?

​LSTM has become one of the most widely used deep learning approaches for time-series modelling and streamflow forecasting. But there is a big difference between using LSTM for rainfall-runoff simulation and using it to make actual forecasts.

​In this live webinar (Link: Watch on YouTube), we explore LSTM for streamflow forecasting—from the fundamentals of the LSTM architecture to increasingly advanced forecasting applications.

​We begin by discussing why LSTM is an option for streamflow forecasting and how its ability to retain information from previous time steps makes it relevant to hydrological forecasting. We then briefly examine the LSTM architecture, including its key components and how it handles temporal dependencies.

​Next, we introduce the forecasting ladder—a framework for understanding the increasing complexity of streamflow forecasting applications, from models that rely primarily on historical hydrological observations to more advanced approaches that incorporate additional information such as upstream conditions and weather forecasts.

​We then walk through several case studies covering a range of LSTM forecasting applications, moving from relatively simple forecasting problems towards more advanced applications, including large-sample forecasting, forecasting in ungauged basins, forecasting using weather forecasts, and other challenging forecasting scenarios.

​The aim is not simply to demonstrate that LSTM can forecast streamflow, but to explore how far we can take LSTM as a practical streamflow forecasting tool.