We are now over a year into the exciting Innowind project funded by Innovation Fund Denmark with some very encouraging results on developing better land surface input for wind modelling with the aim to significantly lower current uncertainties for wind estimates.
At DHI GRAS, we have specifically focused on providing better land cover information and forest height estimates as input to surface roughness models.
Working with our project partners including industry professionals such as Vestas, Vattenfall and EMD International and researchers from the technical university, DTU Space and DTU Wind, is highly valued, as it allows us to streamline our research efforts into something of immediate value for the wind industry.
Our initial results are already showing the value of using satellite-derived data in wind modelling providing improved wind resource assessments.
One of the many key inputs to wind modelers is land cover information as different land cover types can have structural characteristics that affect the wind profile, i.e. forests are more rough than agricultural land etc.
The standard practice by wind modelers is to use generic land cover maps that are not well suited for wind modelling.
An example of this is the Corine land cover that is frequently used by the wind community. This land cover data set has over 40 classes, but most classes are irrelevant for wind modelers and a lot of work goes into merging classes to meaningful groups.
The scale of the land cover map (100m pixel resolution) is furthermore too coarse and although provides good information for regional planning, it often fails to provide finer spatial details required for microscale wind modelling.
Lastly, the Corine data set is only updated every five years. During this time, deforestation and urbanization may have taken place, meaning wind models are often run using outdated land cover data, negatively affecting the models’ performance.
As near real-time satellite imagery and increased computational power are now available, it is about time that the wind industry evolves and improves the quality and resolution of the land cover maps used for wind resource assessments.
The results from a Swedish wind farm show how a few select classes can be converted into input to a surface roughness computation.
Working closely with the project partners, DHI GRAS has developed an automated land cover mapping application that estimates surface roughness via look-up tables (provided by DTU WIND). The model is globally available in 10 meter resolution.
Forest height remains poorly accounted for in the majority of wind simulations and relatively little literature is published on the effect of forest height on wind models.
It is, however, commonly acknowledged that forest height has a major impact on the wind-flow, and not accounting for this will influence wind simulations (we refer to our project partner Ebba Delwik for more information on the subject).
In Innowind, we do the first systematic review of this, and have developed a model to predict forest heights using satellite-based methods and freely available data sources, which are far more cost-efficient compared to traditional LIDAR campaigns. The model requires calibration/validation data of known forest heights, together with predictive variables derived from Sentinel and Landsat satellite sensors.
For forest height modelling, we tested several machine learning regression models. The models take the forest height calibration data and finds relationships between the known heights and the predicted variables from Sentinel and Landsat. The results showed that forest height can be estimated effectively using open source data.
We validated the results by comparing predicted forest height estimations with the validation dataset described above achieving a Mean Absolute Error (MAE) of about 2 meters, which is well within the acceptable range needed for improved wind flow modelling.