How to use Wegaw Snow Water Equivalent data in any hydrological forecast model
In this article
- Why do forecasts matter?
- Snow can be your best predictor
- How to use Wegaw snow observations for short-term and seasonal forecasts
1. Model Timestep
2. Spatial Format - New AI approaches
- Conclusion
Why do forecasts matter?
In regions characterized by snow, like the Swiss Alps, the North American Rockies, or the Scandinavian Mountains, there are clear variations in seasonal streamflow. The accumulated winter snowpack melts during spring and early summer, resulting in a surge of inflow resulting from runoff. However, the intensity appears different each year due to natural variations in mid-latitude atmospheric circulation patterns and local weather systems. Low-snow winters followed by dry summers can result in low water levels in reservoir lakes (when glaciers aren’t present). This was seen in Switzerland after the 2021/22 winter season when Swiss lakes received 20–30% less water from snow melt that resulted in a 23% under-production of hydropower compared to the three previous years.
This annual variability creates a challenge for hydropower providers that sign contracts to deliver energy over multiple years to come (so-called PPAs). Being able to predict spring streamflow into dams early in the year is crucial for implementing timely measures and reducing financial risk. Hydrological models play a key role in providing estimates of expected water volumes throughout the season, aiding in informed decision-making.
As climate change intensifies precipitation variability and diminishes snowfall, future seasonal variations will differ from historical patterns. This underscores the necessity for improved predictive tools to adapt to the changing conditions.
Snow can be your best predictor
The water that arrives in a high-altitude reservoir lake (we call this the inflow or inflow volume) is composed of mainly two elements:
- meltwater from snow (and ice in glacier-influenced areas)
- rainfall
…by observing the snowpack we can estimate the bulk of meltwater arriving over the season, and can then add the range of expected rainfall on top…
In high elevations, meltwater contributions typically surpass those from rainfall. In Switzerland for example, approximately 60% of the water flowing in the Alps comes from snowmelt. As precipitation patterns are different every year and an increasing variability is predicted (and already observed) with climate change, the amount of rainfall is hard to predict months in advance and can only be quantified as a range of potential values.
On the contrary, snowmelt can be estimated months in advance, even before it reaches the reservoir, as it is retained in the snowpack. Hence, the optimal period for predicting the inflow volumes for the upcoming summer is at the end of winter, after the final snowfall. At this time, the observed amount of snow tells us which quantities of meltwater to expect in summer. When saying “amount of snow” we talk about the Snow Water Equivalent (SWE) which is the amount of actual water contained in the snow that becomes available once the snow melts.
So by observing the snowpack we can estimate the bulk of meltwater arriving over the season, and can then add the range of expected rainfall on top, from dry to wet conditions. The models used to do this estimation are called hydrological models and are a common tool of the hydropower industry to forecast inflow volumes to reservoirs. At Wegaw we create entire maps of snow water equivalent every day with unprecedented accuracy that can then be used by such a hydrological model to forecast inflow volumes. If you want to find out how we create these maps, check out the resources on our homepage.
How to use Wegaw snow observations for short-term and seasonal forecasts
Snow observations are often available either only for a single point in time (e.g. drone flight over an area) or a single point in space (in-situ measurement instrument). Data assimilation (DA) methods have been developed to best make use of these observations. Wegaw data is available at any point in space (spatial resolution: 100 m) on every day with a one day latency, which opens new possibilities to integrate this data in different live model setups. The best-suited integration method thereby depends on the timestep and spatial format of the employed hydrological model as discussed in the following. We first focus on process-based models and discuss the integration into AI models separately at the end of this article.
Model Timestep
There are three main ways to use Wegaw snow data in a live forecast setup:
- direct insertion for models running at daily frequency
- data assimilation for models running at higher frequency
- re-calibration
Find a summary of the proposed integration for different model types in Table 1. In the description below we focus on HBV-type process-based models which is currently the most commonly-used model. Get in touch with our team here to discuss integration methods for other model types.
1. Direct Insertion
→ Use for any model running at daily frequency
As Wegaw SWE maps are available on a daily basis, they can be a direct input for hydrological models that run at the same daily frequency. Keeping in mind that the Wegaw SWE comprises the solid AND the liquid water content of the snowpack, certain adaptations to the snow module are necessary if the model does not contain an explicit SWE state. (If the model does contain an explicit SWE state, it can simply be replaced by Wegaw SWE.)
In the case of a single-layer snow module with one solid (ice) and one liquid model state (LWC=liquid water content) we recommend to continue modeling of the LWC as usual and to only adapt the solid state. The solid state will then be updated by the Wegaw SWE and LWC
and the LWC is updated using the previous solid state and meteorological inputs.
For HBV-type multi-layer snow modules the snow layers can be scaled using the Wegaw SWE after applying the processes implemented in the module. Again, we recommend only updating the solid part of the snow state.
Optionally, the rain/snow separation of precipitation can be adapted in the following way:
A potential transfer function might be necessary to scale the Wegaw SWE to the simulated SWE-state which are often of different amplitudes. We recommend instead to re-calibrate the snow module parameters of the model when using direct insertion of Wegaw SWE to avoid the need for a transfer function.
2. Data Assimilation
→ Use for any model running at sub-daily frequency or highly parameterized daily models
Data assimilation methods are used to integrate observation data that is not available at the model frequency such as snow survey results. The model’s snow state is thereby corrected by using current observations and their uncertainty. Inserting observations directly into higher-frequency models can lead to unstable results; thus the use of common data assimilation methods is advised, e.g. Particle filter, Kalman filter, variational assimilation.
An alternative to these methods and a very simple data assimilation strategy is to work with relative changes. The model’s snow state can be adjusted be applying the same relative change as observed in the Wegaw SWE time series over a specified number of n time steps:
We recommend to apply the change only to the solid part of SWE (ice):
This method is much easier to implement but might lead to instabilities in the predicted time series. However, this can be reduced by re-calibrating the model parameters as described in point 3.
3. Re-calibration
In combination with either of the two integration ways described above, the model parameters (either full model or only snow module) can be re-calibrated to
- avoid transfer function and/or
- improve snow-related model parameters (snow/rain separation, melting factor, etc.).
For direct insertion, the model is re-calibrated by minimizing the error on inflow simulation when inserting Wegaw SWE time series as described in point 1.
When using data-assimilation a synthetic SWE value will still be generated by the model. For model stability it can be helpful to re-calibrate the parameters of the snow module to best fit the observed Wegaw SWE. The optimization criteria is thus to minimize the error of the modeled SWE with respect to Wegaw SWE as the target.
Table 1: Different types of hydrological models and proposed integration method of Wegaw SWE data.
Spatial Format
Models differ in their spatial format. From lumped to semi-distributed to distributed models the input data is either averaged over the entire area of interest (AOI), over sub-divisions of the AOI or treated on a grid-base.
We observe a general trend towards spatially distributed models that work on grid-based inputs. The main advantage is that the model then works in the format of the input data (mostly maps of meteorological variables) which enables to pick up on small-scale effects and simplifies the integration of satellite data.
We observe a general trend towards spatially distributed models to utilize satellite data.
Wegaw SWE maps fit all of the different spatial formats as we provide the data to our clients directly in their desired format, either grid-based as geotiffs or netcdf or zonal averages as csv (further description in Table 2). The historical SWE maps, which we always generate for several years in the past for quality checks, further allow us to create a statistical analysis of the snowpack in your area, to propose the best possible sub-division of the AOI if desired.
Table 2: Spatial formats of hydrological models and corresponding Wegaw data delivery format.
New AI approaches
Hydrological models based on artificial intelligence are increasingly applied for inflow forecasting. In a research setup they have proven to outperform conceptual models with the ability to simplify the integration of new data sources, such as Wegaw SWE maps. Commonly used model types are Graph Neural Networks, Support Vector Machines and Long Short-Term Memory Networks. To predict a time series the model is often split into two parts, a hindcast module and a forecast module, to allow for different inputs before and after the forecast point.
Wegaw data is particularly suited for forecasts based on AI methods as it is generated with our unique AI processing pipeline. The data can be simply added to the list of inputs of the hindcast module if it is running at daily frequency. If you require SWE data at higher frequency, get in touch with our team here to discuss our available temporal interpolation methods.
Conclusion
By integrating Wegaw’s high-resolution SWE maps into hydrological forecasts, we have demonstrated in several validation setups the improvements of the forecast accuracy when using daily snow maps. As elaborated in depth in this article, the data we produce is compatible with all different kinds of hydrological model setups and we customize the delivered live stream to fit your individual requirements.
While consistently working on further improving the quality of our data, we are developing a cost-effective solution of a simplified seasonal snow- and glacier melt estimation to support asset management without hydrological modeling, that will be launched soon. Reach out to our team if you want to discuss customized solutions for your business.
To learn about how Wegaw snow data can be used for solar power, check out our latest article on Super-Resolution Snow Height Mapping.
References
Zappa M. et al. 2012: Klimaänderung und natürlicher Wasserhaushalt der Grosseinzugsgebiete der Schweiz
SLF, Winterbericht 2021/22
BAFU, Hydrologisches Jahrbuch der Schweiz 2022
