The Long Short-Term Memory (LSTM) model is a deep learning method that has proven very competitive with regard to streamflow predictions over recent years. In contrast with basic recurrent neural networks, the LSTM’s memory line allows the model to retain information over long periods of time, solving the vanishing gradient problem. In this work, we applied the LSTM model in the region of the Great Lakes, first in a “hindcast” mode when fed with dynamic forcings coming from the Canadian Surface Reanalysis version 3.2 (CaSR v3.2) produced at Environment and Climate Change Canada (ECCC), and then in a “forecasting” mode when simply replacing the last days of the 365-days lookback window of the data cube provided to the LSTM model with actual forecasts from ECCC’s Global Deterministic Prediction System (GDPS). To do so, the model was first trained over 2001-2018 and with a set of 212 streamflow gauges in the Great-Lakes region, and tested over 2019-2023, when using only CaSR 3.2 as the source of the dynamic forcings needed by the model to evaluate its temporal robustness. Then, the model was applied over a full hydrologic year spanning over 2021/2022 in a forecast mode, producing streamflow forecasts up to 6 days. These LSTM forecasts are compared to the streamflow forecasts that were performed with the National Surface and River Prediction System, which consists of ECCC’s distributed and physically-based hydrologic forecasting system. While the LSTM already shows very promising results when compared to NSRPS forecasts, there is still room for improving the LSTM forecasts and combining the strengths of both systems, as well as further work needed to prepare the LSTM model for operational deployment at ECCC.

Accurate estimation of percolation is crucial for assessing landfill final cover effectiveness, designing leachate collection/treatment systems, and many other applications, such as in agriculture. Despite the importance, percolation is seldom measured due to the high cost and maintenance of lysimeters, underlining the need for skillful simulation. Process-based numerical models, despite requiring validation and numerous parameters, present an alternative for percolation simulation, though few studies have assessed their performance. This study compares percolation measured from three fully instrumented large-scale experimental plots to simulated percolation using a new version of the Soil Vegetation and Snow (SVS) land-surface model with an active soil-freezing module. Previous research indicates numerical model performance may significantly vary based on soil-related parameter values. To account for input data and parameter uncertainty, we use an ensemble simulation strategy incorporating random perturbations. The results suggest that SVS can accurately capture the seasonal patterns of percolation, including significant events during snowmelts in spring and fall, with little to no percolation in winter and summer. The continuous ranked probability skill score values for the three plots are 0.13, -0.13, and 0.33. SVS simulates near-surface soil temperature dynamics effectively (R 2 values 0.97-0.98) but underestimates temperature and has limitations in simulating soil temperature in snow-free situations in the cold season. It also overestimates soil freezing duration, revealing discrepancies in the onset and end of freezing periods compared to observed data. This study highlights the potential of land surface models for the simulation of percolation, with potential applications in the design of systems such as leachate collection and treatment. While the SVS model already provides an interesting outlook, further research is needed to address its limitations in simulating soil temperature dynamics during soil freezing periods.