HydroClimateSight Feature Highlight: Understanding Ensemble vs Deterministic Forecasting in HydroSphereAI

New in HydroSphereAI: Short-Range Forecasting

We’re excited to announce that short-range streamflow forecasting is now available in HydroSphereAI! Our new short-range forecasts offer detailed predictions for the day ahead with hourly output intervals. To address the lack of uncertainty information typical of deterministic systems, we’ve introduced a lagged ensemble approach, giving you a clearer picture of forecast confidence in the near term. In this post, we’ll walk through how deterministic and ensemble forecasts differ, how short-range and medium range forecasts are used in HydroSphereAI, and what the new short-range functionality means for your water management decisions.

Ensemble vs. Deterministic Forecasting: Understanding the Differences

In hydrological forecasting, the distinction between deterministic and ensemble forecasts is a critical one, especially when it comes to understanding uncertainty, timing, and confidence in streamflow predictions. Our HydroSphereAI platform provides both short-range and medium-range forecasts for streamflow prediction across Canadian watersheds. Each range is powered by a different type of meteorological forecast model, and it's important to understand how to interpret each correctly.

Figure 1. Deterministic forecast (using HRDPS) at Station: 02HB013 (CREDIT RIVER NEAR ORANGEVILLE), illustrating a single possible outcome with no indication of uncertainty. Note: Forecast outputs in HSAI are presented differently than shown here.

What’s the Difference Between Deterministic and Ensemble Forecasts?

Deterministic Forecasts: Applied to weather prediction, a deterministic forecast is a single forecast generated from one “best estimate” of the current state of the atmosphere. It shows one possible future, assuming we know the current weather conditions exactly, which, of course, we never do. The High Resolution Deterministic Prediction System (HRDPS), is a prime example. It provides a single, high-resolution weather prediction every 6 hours. In HSAI this translates to a single streamflow forecast for each issue time. (see Figure: 1)

Figure 2. Ensemble forecast (using GEPS), showing multiple forecast traces generated from slightly different initial conditions at Station: 02GG003 (SYDENHAM RIVER AT FLORENCE), illustrating a range of possible future outcomes and their spread to capture uncertainty.

Ensemble Forecasts: In contrast, ensemble forecasting runs the same weather prediction model many times (often 20+), each with slightly different initial conditions to reflect uncertainty in the atmosphere. The result is a range of possible scenarios. HydroSphereAI’s medium-range forecasts are powered by an ensemble of 21 meteorological predictions supplied by Environment Canada’s Global Ensemble Predictions System (GEPS), allowing us to visualize not just what might happen, but how likely different scenarios are. (see Figure 2).

Why Ensemble Forecasts Matter

With ensemble forecasts, you can begin to think probabilistically. Rather than asking, “What will the flow be next Tuesday?” the better question becomes: “What is the range of possible flows, and how likely are we to exceed a certain threshold?”

For example:

• A wide spread in the ensemble indicates high uncertainty, with many different possible outcomes.

• A narrow spread means the ensemble members are agreeing, and confidence is higher.

Figure 3. A lagged ensemble forecast at Station: 02AB014 (NORTH CURRENT RIVER NEAR THUNDER BAY). Four recent HRDPS forecast runs are combined, with each run further diversified by multiple machine learning models. This approach generates a spread of short-term streamflow predictions that reflect both recent model trends and uncertainty.

Also keep in mind that the median line of the ensemble is not an actual trajectory: a long, broad peak may be the result of many distinct, sharp peaks that occur with some temporal shift (timing uncertainty) – in this case the uncertainty range around the broad peak would also be large

Understanding this helps users prepare for a range of possibilities rather than anchoring to a single, potentially misleading forecast.

Bridging the Gap: Lagged Ensemble for Short-Range Forecasts

Because short-range deterministic forecasts don’t offer a measure of uncertainty, we’ve developed a lagged ensemble approach in order to provide a measure of forecast uncertainty.

Here’s how it works:

• We take the last four HRDPS forecast runs (from the past 24 hours) and aggregate them into a mini-ensemble, creating a spread of possible streamflow scenarios.

• In addition, each HRDPS forecast is run with 10 slightly different ML models, further increasing the ensemble spread

• This allows us to show an uncertainty range for short-term forecasts, based on recent model trends. (See Figure 3).

Figure 4. Example of a lagged ensemble combining four consecutive forecasts. This figure shows four individual deterministic forecasts issued six hours apart, each starting and ending slightly later than the previous one. When these four forecasts are combined into a lagged ensemble, the result is a single, broader distribution that captures the spread between issue times. The overall range of the lagged ensemble is wider than the spread of the four forecasts alone because each of those forecasts represents the median of an underlying parameter ensemble, so the combined lagged ensemble also incorporates the variability from those ensembles.

However, there is a trade-off: since one of the forecasts is up to 24 hours old, this shortens the usable forecast horizon by a day. Effectively, it provides a 1-day uncertainty window, rather than the full 2-day deterministic forecast (Figure 4)

We’re also exploring the integration of U.S.-based ensemble data, such as HRRR, to expand the number of ensemble members and further improve short-term uncertainty estimation.

HydroSphereAI: How We Use a 4-Member Lagged Ensemble Approach vs. Ensemble Forecasts

In HydroSphereAI:

• Short-range forecasts use a 4-member lagged ensemble, built from the last four HRDPS forecast runs over the past 24 hours. By aggregating these runs into a mini-ensemble, we generate a range of possible streamflow scenarios, allowing us to present an uncertainty range that reflects recent model trends.

• Medium-range forecasts (2–16 days) use ensemble weather data from 21 different forecast members, allowing us to present a probabilistic forecast that captures the range of potential streamflow scenarios.

What This Means for You

Whether you’re monitoring an approaching flood or planning operations around a potential low-flow event, it’s crucial to interpret forecasts correctly:

• Use the short-range ensemble for immediate planning and risk assessment, especially where timing and detail are key, but keep in mind that the ensemble size is small and not all possible scenarios may be captured

• Use ensemble forecasts to assess risk, uncertainty, and confidence in the medium range.

• Understand the spread: A flat or broad median curve with a wider probability distribution doesn’t necessarily mean low flows, it usually reflects high uncertainty in the timing of sharper peaks in different scenarios.

We believe this dual approach offers the best of both worlds, precision in the short term and meaningful probabilistic guidance in the medium term.

We Want Your Feedback

We’d love to hear from you, especially about the new lagged ensemble for short-range forecasts. Do you find it helpful? Would you prefer to see the individual forecasts instead, maybe as a “spaghetti plot” with multiple forecasts in one graph?

Please reach out and let us know how HydroSphereAI can better serve your forecasting needs. Send us an email at HydroClimateSight@aquanty.com to learn more, or schedule a demo with one of our product experts.