The use of Gaussian process regression for wind forecasting in the UK

Wind energy has experienced remarkable growth in recent years, both globally and
in the UK. As a low carbon source of electricity this progress has been, and
continues to be, encouraged through legally binding targets and government policy.
However, wind energy is non-dispatchable and difficult to predict in advance. In
order to support continued development in the wind industry, increasingly accurate
prediction techniques are sought to provide forecasts of wind speed and power
output.
This thesis develops and tests a hybrid numerical weather prediction (NWP) and
Gaussian process regression (GPR) model for the prediction of wind speed and
power output from 3 hours to 72 hours in advance and considers the impact of
incorporating atmospheric stability in the prediction model. In addition to this, the
validity of the model as a probabilistic technique for wind power output forecasting
is tested and the economic value of a forecast in the UK electricity market is
discussed.
To begin with, the hybrid NWP and GPR model is developed and tested for
prediction of 10 m wind speeds at 15 sites across the UK and hub height wind
speeds at 1 site. Atmospheric stability is incorporated in the prediction model first
by subdividing input data by Pasquill-Gifford-Turner (PGT) stability class, and then
by using the predicted Obukhov length stability parameter as an input in the model.
The model is developed further to provide wind power output predictions, both for a
single turbine and for 22 wind farms distributed across the UK. This shows that the
hybrid NWP and GPR model provide good predictions for wind power output in
comparison to other methods. The hybrid NWP and GPR model for the prediction of
near-surface wind speeds leads to a reduction in mean absolute percentage error
(MAPE) of approximately 2% in comparison to the Met office NWP model.
Furthermore, the use of the Obukhov length stability parameter as an input reduces
wind power prediction errors in comparison to the same model without this
parameter for the single turbine and for offshore wind farms but not for onshore
wind farms. The inclusion of the Obukhov length stability parameter in the hub
height wind speed prediction model leads to a reduction in MAPE of between 2 and
iv
5%, dependent on the forecast horizon, over the model where Obukhov length is
omitted. For the prediction of wind power at offshore wind farms, the inclusion of
the Obukhov length stability parameter in the hybrid NWP and GPR model leads to
a reduction in normalised mean absolute error (NMAE) of between 0.5 and 2%. The
performance of the hybrid NWP and GPR model is also evaluated from a
probabilistic perspective, with a particular focus on the appropriate likelihood
function for the GPR model. The results suggest that using a beta likelihood function
in the hybrid model for wind power prediction leads to better probabilistic
predictions than implementing the same model with a Gaussian likelihood function.
The results suggest an improvement of approximately 1% in continuous ranked
probability score (CRPS) when the beta likelihood function is used rather than the
Gaussian likelihood function.
After considering new techniques for the prediction of wind speed and power
output, the final chapter in this thesis considers the economic benefit of
implementing a forecast. The economic value of a wind power forecast is evaluated
from the perspective of a wind generator participating in the UK electricity market.
The impact of forecast accuracy and the change from a dual imbalance price to a
single imbalance price is investigated. The results show that a reduction in random
error in a wind power forecast does not have a large impact on the average price per
MWh generated. However, it has a more significant impact on the variation in price
received on an hourly basis. When the systematic bias in a forecast was zero, a
forecast with NMAE of 20% of capacity results in less than £0.05 deviation in mean
price per MWh in comparison with a perfect forecast. However, the same forecast
leads to an increase in standard deviation of up to £21/MWh. This indicates that
whilst a reduction in random error in a forecast might not lead to an improvement in
mean price per MWh, it can lead to a more stable income stream. In addition to this,
Chapter 6 considers the use of the probabilistic and deterministic forecasts
developed throughout this thesis to choose an appropriate value to bid in the UK
electricity market. This shows that using a probabilistic forecast can limit a
generator’s exposure to variable prices and decrease the standard deviation in hourly prices.