WC146 JanFeb 2026 - Magazine - Page 27
models. Areas that present unique challenges include:
THE PRAIRIES: Forecasting here is notoriously difficult due to
factors like localized storms, data quality issues and the fact
that runoff is a tiny fraction of the overall water balance.
URBAN WATERSHEDS: The complication here isn’t the natural hydrology—it’s human interference, such as with dams and flow
regulation, which makes it tougher to forecast without access
to operational data.
It’s vital to note, however, that while HSAI can forecast without
a physical gauge, the model wouldn’t exist without decades of realworld data collection, including the input of ongoing information
from existing gauges. Aquanty is clear that the technology is
designed to augment existing flood forecasting programs and not
replace the need for continued investment in streamflow gauges and
data acquisition. The skill level of the technology is a testament to
the decades of data work done by government agencies that allow
this kind of innovation to flourish.
Predicting the predictors
HSAI’s advanced skill is built upon
decades of historical data, but its daily
performance relies on current weather
predictions. Every day, Environment
and Climate Change Canada (ECCC)
publishes multiple forward-looking
forecast products—some with as many as 21 separate forecasts per
day. These forecasts are based on subtle, random perturbations in
current atmospheric conditions, and they capture the probability
of how the weather might actually play out over the next two
weeks.
HSAI runs its algorithm against every single one of these
forecasts. By pairing its robust hydrological modeling with
this ensemble of atmospheric possibilities, the system is able
to generate a clear picture of the possible range of streamflow
outcomes, predicting the maximum, minimum and median
expected streamflow.
Naturally, the accuracy of any forecast decreases as the time
horizon extends. The models have a much better chance of
accurately predicting streamflow tomorrow versus 14 days from
now. In practice, the best streamflow forecasts are typically
delivered two to three days ahead of an actual event. However, by
using the full two-week forecast horizon, forecasters gain crucial
lead time to assess risk and prepare for possible high-impact
scenarios.
“By learning from this wide range of data, the model can generalize
its knowledge, meaning it can forecast stream昀氀ow in areas that don’t
have physical gauges.”
“One of the challenges with urban areas is the fact that
they’re highly controlled reservoirs… Dam operators will
release water based on how the reservoir is behaving on a
given day, not on a schedule,” explains McNeill. “That type of
behaviour is not captured in the inputs to the algorithm… but
it would be very easy to come by with the right partners.”
McNeill also emphasizes that due to the nature of machine
learning algorithms improving as new data becomes available,
Aquanty expects to see improved performance year over year in
all areas, including in urban catchments.
A game-changer for local flood risk
The ability of HSAI to generalize its knowledge offers
significant practical benefits, particularly for smaller
municipalities and conservation authorities (CAs) that often
operate with limited resources.
For many years, streamflow forecasting has been anchored to
the physical location of a streamflow gauge. Building a custom,
high-fidelity model for a new location traditionally costs tens
of thousands of dollars and requires extensive, labour-intensive
calibration. This means that smaller CAs may lack the funding
to build their own local models.
HSAI offers a scalable way to deliver detailed forecasts
beyond the fixed hydrometric network.
“We can drop a pin on a map anywhere along a river
anywhere across Canada and we can automatically delineate
that upstream area. Then, we can start predicting streamflow
at that location with or without the presence of an actual
streamflow gauge,” explains McNeill.
Crucially, this cost-effectiveness doesn’t compromise
accuracy. The ability to deploy accurate forecasts almost
anywhere in Canada quickly is what makes the technology a
useful complementary tool for jurisdictions from small, remote
communities to mid-sized cities.
WAT E R C A N A D A . N E T
Pushing the boundaries of streamflow forecasting
HSAI represents a significant development in hydrological
forecasting—expanding the accessibility and reach of detailed
predictions. By leveraging a massive continental dataset and
the advanced Long Short-Term Memory (LSTM) architecture,
the technology demonstrates the ability to generalize its
knowledge. As the sector grapples with the uncertainties of a
changing climate and the constraints of existing infrastructure,
tools like HSAI complement Canada’s established flood
mitigation efforts by turning extensive data into operational
insights. The technology confirms that continued investment
in both real-world data collection and advanced ML
integration is key to advancing the resilience of Canada’s water
management systems.
WATER C AN ADA • JANUARY/ FEBRUARY 2026
27
