How much energy will my favourite wind farm produce tomorrow? I have often wondered if it was possible to set up a DIY forecasting tool for offshore wind, and in this post I’ll briefly describe one possible way to do this with publically available datasets.
The first thing to do is to gather wind speed and direction forecasts, and for this analysis I have used the Meteo France data, kindly made available in a nice and tidy format by @OpenMeteo at https://mf-models-on-aws.org/. I have chosen the arpege-Europe dataset, which provides hourly forecasts up to 102h at approx. 11 km resolution; see the computational domain below:
Then, I picked one of my favourite wind farms: Westermost Rough, located near the East Coast of the United Kingdom. It consists of 35 6MW-154m turbines, and has been operating without major issues so far. Half-hourly , metered power time series for this wind farm are available from the Elexon BMRS B1610 database, see this previous post for more details.
The last thing to do was to compute a park power curve (including only wake effects, and not other losses) for every wind direction; let me warn you that this is a very simplistic way of handling park power curves (in particular the directional averaging, which I have not carried out here – see also this previous post for more info on how to model electrical and availability losses). With this park power curve, I can turn every {speed;direction} values at ~hub height (100 m) into power.
I have downloaded forecast data between the 29th of September and the 3rd of November, see below an example of forecasted time series against the actual wind farm power: the forecast issue date is marked with a red dot, the forecasted wind speed and power are in yellow, and the actual wind farm power in black.
Now here is the same forecast, followed by the next five forecasts:
And here is another illustrative example:
And… here are the stats: I have cheated an only selected relative errors smaller than +/- 2.. , and still I am getting very large rmse values (about twice larger than what would you should expect from the literature for a good forecast).
Also, surprinsgly enough, there seem to be no dependency to the time horizon.. I really need to look into this…
Still, I’m happy I could put together such a simple forecasting tool based on publicly available data ^^. Maybe are these data useful for other purposes, including validation work for forecasting models?
Comments/questions are welcome,
Rémi
