We use machine learning with advanced neural networks to accurately forecast production. There is big data processing, a data hub that we built ourselves and is actually quite cool, and even a data lake for complete buzzwordyness. And once we have the forecast, we start and stop the machines to optimise production. Hard problems = good times!
5. How windmills work
o Wind turbines convert kinetic
energy in the wind into
mechanical power.
o Generators in the turbine
nacelle converts this mechanical
power into electricity
7. o Wind farms must nominate expected power output to the
grid operator (ENTSOE) the day before
o Operator asks for more or less until expected consumption
has been nominated.
o Operator can order shut-down (curtailment)
• Too much power in the grid = busted electronics
• Not enough power in grid = brownouts/blackouts
How the market works - nominations
8. o Production is priced per 15 minute unit
o … price fluctuates constantly (think stock market)
o ... the maximum deviation from 0 within the time unit is the
price for the whole unit
• 11:00 125 Euro/MWh.
• 11:08 140 Euro/MWh.
• 11:12 -141 Euro/MWh.
Imbalance market
13. Numerical weather
predictions
o Weather forecasts are generated by
numerical weather prediction models
o NWP models uses mathematical models
of the atmosphere and oceans to
predict the weather based on weather
data assimilated at model run time
o NWP models are solved numerically
using finite difference schemes in a
predefined grid.
15. Inaccuracies
o Imperfect modelling of the atmosphere
o Imperfect parameterization of topology
o Imperfect data assimilation (now-
weather)
o Sub-grid scale weather phenomena
(ultralocal)
o Temporal resolution (time)
16. Correcting the NWP:
Static conditions
o Static factors - dirty wind, topology =
systematic bias.
o Simple statistical corrections for
systemic bias
o Site-specific error correction of NWP
17. Correcting the NWP – measured conditions
o Sensors measure actual wind:
• Direction
• Speed
• Gustiness
o Other variable factors – humidity,
cloud or sun etc.
o Sensor data is cleaned using
simple learning algorithms
• Outlier detection
• Identify operational state
• Temporal & grid interpolation /
aggregation
30. Birds
Dear all,
To day the young sea eagle has leave the nest.
DTbird on Turbine 28 is unfortunately not yet in place.
We decide to stop T29, T30 and T31 to protect the bird.
Best regards,
Niek
31. o Sometimes, flocks of
bats fly through the
area.
o We are looking at
training models to
predict when.
o And send a signal to
shut down.
Bats
…yes, the Bat Signal.