Ensemble Forecasts for Extreme Weather Early Warning Systems
On October 27th, the most severe storm in over the past 50 years struck South-Eastern Estonia. Falling trees, flying roofs, and severe winds wreaked havoc over the area. Electricity was lost in more than 60,000 households, impacting the vast majority of this coarsely populated area. For many, the ensuing power outages lasted for over three days.
After the storm, residents in high-risk areas complained that advisories and warnings before the storm were either unclear or nonexistent.
Seeking ‘HARMONIE’ in the midst of a Storm
With the topic of storm advisories and accessible storm forecasting in mind, we are highlighting the meteorological model HARMONIE. This groundbreaking, holistic model is now available on the Planet OS Datahub. Using HARMONIE, we take a look at the ensemble forecast data from this weather model in areas of Estonia most impacted by the devastating storms.
(If you would like to learn how to use the MetNO HARMONIE Ensemble forecast yourself and create visualizations like the ones below, check out our Jupyter notebook that provides a step-by-step tutorial.)
Provided by the MEPS, HARMONIE exists as a cooperation project between the Norwegian, Swedish, and Finnish meteorological institute. Referred to as AROME in France, it is a high-resolution, short-range weather prediction model. HARMONIE is used in several Central European and Scandinavian countries. In the MEPS configuration, it has 2.5km grid cell size with capabilities at even higher resolutions, such as a 1.1km resolution throughout France.
In addition to being a high-resolution model, HARMONIE is also implemented as an ensemble. Ensemble implementation means that for each forecast time the weather model produces ten separate simulations. The first one is referred to as a control member or deterministic member, while the other nine are ensemble members. Each ensemble member differs from the other by small changes in the model’s initial conditions. The ensembles help in estimating the variability of weather. For example, if each of the ensemble members show a similar forecast, the accuracy of the forecast is expected to be higher. While the concept of ensemble prediction has been used globally, and 10-day weather forecasts for decades, it is a relatively new and recently applied strategy in short-range forecasting.
We can see HARMONIE in action with the example below. We will observe meteorological observations from the national observation network.
Storm Wind Speeds in Estonia
Notable measurements on the above map of Estonia are those on the lower right corner, where the observed 26 m/s wind speeds are higher than any recorded windspeeds in the past 50 years.
From Mapping to Forecasting
The first forecast we look at has an initial time of 2019–10–25T12, which in Estonian local time becomes available after 4:30 PM. On the graph below, the spatial distribution of maximum wind speed over the course of the day is plotted.
Wind Gust Forecasts
Through the map above, it is evident that there will be strong winds on the western coast and surrounding islands, as well as in the central and southern parts of the country. In fact, it was found that both the distribution of strong wind areas and maximum values displayed by the forecast were impeccably similar to real-time observations.
So, do we have a perfect forecast that can be used to issue accurate weather warnings regionally, or is it too good to be true? In order to discern the true effectiveness and accuracy of the HARMONIE forecasts, let’s take a look at the subsequent forecasts.
In the next graphs below, we created supplementary forecasts that were issued every six hours and for the exact same time period (the day of October 27th, 2019). It should also be noted that as the forecast lead times decrease in duration, they typically become more accurate.
Cross Comparing Forecasts
The first trends we can observe from these forecasts compared to the previous forecasts are weaker wind speeds both over the seas and inland. However, all forecasts do show maximum wind speeds over 20 m/s in the South-Eastern region of Estonia. Friday evening and Saturday morning forecasts leave the impression that the storm is weakening. With this we wonder; how should one interpret a situation where subsequent forecasts have such drastic differences?
With this considered, we will now take a look at the ensemble forecast, rather than the main deterministic forecasts alone.
Ensemble Forecasts in Action
In the next graph, we plotted the median values of the HARMONIE ensemble. It becomes evident that the differences between subsequent forecasts are not as drastic as when we look at the control member alone. Similarly to the deterministic forecast, we see that wind speeds are lower, especially over the sea. However, strong winds in South-Eastern Estonia do remain.
The next question we might ask is whether the ensemble median is the best way to express the statistical information that the ensemble itself contains. So, let’s look at the ensemble forecasts with actual distribution given for weather stations. We also display actual observed values and deterministic forecasts on the following boxplots as well.
Maximum Wind Speed Ensemble Forecast
Boxplots show the spread of ensemble forecasts for different weather stations:
Due to their proximity to the locations that suffered the most damage in the October storms, we are most interested in the Valga and Võru stations. We can see that on all of the three forecast lead times, the ensemble spread is small for these stations and that the ensemble median is always above the 20 m/s value.
When looking at the deterministic forecast only, it may appear that a strong wind signal is lost from the forecast after 2019–10–25T12. But, when looking at the ensemble we still observe a high probability for strong winds. Furthermore, it is typically not a good idea to draw conclusions about a weather forecast based on a single case alone.
This current use case illustrates how using the ensemble can make our assessment of the forecast reliability more robust.
To tie this weather event together, we have created a few more graphs. Below we can see the ensemble forecasts for station Võru (the lower right on the map) from different forecast times.
A Closer Look: Ensemble Forecasts in Võru
In each of the forecasts, the median value of the maximum wind speed was above 20 m/s. It is also apparent that the deterministic forecasts show a sharper decline in the Friday evening and Saturday morning forecasts, leading to the possibility of false storm weakening signals.
The next graphs are all ten ensemble members for both 2019–10–25T12 and 2019–10–26T00 forecast times. For these two forecasts, we observed a large difference in both deterministic and ensemble median values. But, how significant are the differences when we look at the forecasts at the level of individual members? Are the differences negligible or is it a vastly different weather forecast?
For 2019–10–25T12 we see that only one member out of ten projects weak winds over Estonia. Seven members illustrate strong winds, measuring out at over 24 m/s.
For 2019–10–26T00 we see that weak winds over land prevail in four ensemble members and strong winds also in four members. Overall, six ensemble members predict stronger than 20 m/s winds over Estonia’s landmass, which is still a significant probability. The common features for these two initial times are significantly different forecasts within the same ensemble and ensemble members with high wind speeds. This indicates that in the case of complex weather phenomena, it is useful to look at the individual ensemble members in addition to the forecast as a whole.
Takeaways
We have shown that in the case of extreme weather situations, using ensemble forecast systems can improve our understanding of the probabilities of emergency weather patterns. However, translating the probabilities shown in forecasts into weather advisories that are distributed to the public is an integral step for proactive disaster preparedness and response. These efforts are far from trivial and are important for raising awareness. Although, work is still left to be done in creating efficient mechanisms for integrating these forecasts into structured responses.
As seen above, HARMONIE can be an incredibly useful resource to provide a multi-faceted forecast. At Planet OS, we are motivated to make game-changing resources that are used to create these forecasts as accessible and easy to use as possible. As data on the nearly infinite patterns and changes in Earth’s climate and environment increases in quantity and distribution, it is important we equip ourselves with this information. Climate and environmental data and computing can hold the key to advancing current technologies and responses to extreme weather events.
Have a climate story or related publication? Let us know and we can feature your work in our blog or newsletter.
If you have any questions or requests for additional datasets, you can join our Planet OS Slack community or reach out via email.
For more examples of how you can use the Planet OS Datahub API to work with high-quality weather and environmental data, check out our Jupyter Notebooks or visit data.planetos.com. If you like to receive email updates when new data becomes available, subscribe to the Planet OS newsletter.
