One of the major challenges in forecasting the power generated by almost fully weather-dependent renewable energy sources such as wind and PV is the inaccuracy of the employed weather forecasts (numerical weather predictions – NWPs). These forecasts – especially the forecast horizon of 0-4 h – are essential for the energy market. Uncertainties and forecast errors reduce the profits of direct marketers or even lead to losses, while affecting the grid stability among grid service providers.
On the way towards a self-learning forecasting system, the forecasting quality is to be improved further by testing combination methods of existing forecasts from different NWPs and (re)calibrating them in real time. To this end, an additional model (frequent updates, by IBM) is used and its benefit for the combination is determined.
The project is interesting to:
Direct marketers, grid operators, utility companies
Investigating the benefit of the IBM GRAF weather model blend for our wind power forecasting
Developing a framework conforming to continuous delivery for machine learning (CD4ML) to combine different wind power forecasts using different methods
Selecting and optimizing appropriate combination methods
Improving the forecasting quality
Detailed project description
It is known that a simple combination of power forecasts from different NWPs (numerical weather predictions) with statistical weights already increases the forecasting quality and significantly reduces errors. If the combination is continuously adapted during operation, the forecasting quality may be improved even further, since weather models vary in their accuracy of mapping different weather situations. The combination in the GRAF-CoWi project is planned to include power forecasts based on already existing weather models and additionally the IBM GRAF.
In addition, methods are compared to answer the following questions:
Does an adaptive combination offer better performance than static combination?
Are non-linear methods superior to linear methods?
To this end, a framework is set up that provides the possibility of automatically validating different combination methods against each other and selecting the optimal method for the respective situation. New approaches can be easily and automatically integrated by changing configurations, simulated for past periods, and compared with each other, making it possible to quickly analyze and evaluate the benefits of a newly developed method. The same applies when adding new weather models – adaptive methods do not require new static weights to be calculated over extensive historical periods (that are initially not available).
The combination model for wind power forecasting is freely configurable in terms of recalibration frequency and past training period, among other parameters. That is why the module is used directly during operation for this purpose, but is not limited to the combination of wind power forecasts.
Project participants: Alina Herzog, Dr. Axel Braun, Jonas Koch, Jens Hoppe, Vitalij Pankraz, Wolfgang Slaby
Direct marketer Enercity and the separately operated virtual power plant (EnercityVPP) supported the researchers in particular by providing data and their professional assessment from the user’s perspective.