Environmental Engineering Reference
In-Depth Information
approach are replaced by a direct conversion from the input data, that is, the NWP
forecast and the online wind power measurement, to a wind power forecast.
A simple statistical model can be based on the following form (Kariniotakis
et al. , 2006b):
P
t NWP Þ; qð
ð
t
þ
k
j
t
Þ¼
f
ð
P
ð
t
Þ; u
ð
t
þ
k
j
t
þ
k
j
t NWP Þ; x
ð
t
þ
k
j
t NMP ÞÞ
P
ð
t
þ
k
j
t
Þ
is the power forecast for time t
þ
k made at time t , P ( t ) is the power
measurement at time t ,
is the NWP wind speed forecast for time
t þ k made at time t NWP (the time of the last NWP run), t þ k j t NWP Þ is the NWP
wind direction forecast for time t
u
ð
t
þ
k
j
t NWP Þ
^
þ
k made at time t NWP and x
ð
t
þ
k
j
t NWP Þ
is
another meteorological variable forecast by the NWP model for time t
k made
at time t NWP . The function f can be linear, e.g. ARMA (autoregressive moving
average) or ARX (autoregressive with exogenous variables), or it could be non-
linear, e.g. NARX (non-linear autoregressive with exogenous variables), NN
(neural network) or F-NN (fuzzy neural network). These statistical models require
extensive use of training sets of forecast and measured data to capture the spatial
and temporal dependencies in the time series data, by identifying the model
parameters and functions needed to reproduce the relationships between the input
explanatory variables and the forecast wind power.
The wind power prediction tool (WPPT) developed by Informatics and
Mathematical Modelling (IMM) at the Danish Technical University (DTU) is an
example of a wind power forecasting system which uses the statistical approach to
the downscaling and wind to power forecast tasks. A schematic of the WPPT sys-
tem is shown in Figure 6.20. WPPT has been operational in western Denmark since
1994 and uses statistical non-parametric adaptive models for prediction of power
from selected wind farms. Data checking is applied to the online input power
measurements from the selected wind farms. It also uses statistical up-scaling to the
full installed capacity in the region or sub-region.
A typical result taken from Giebel et al. (2003) for the application of WPPT to
the western part of Denmark is shown in Figure 6.21, showing a comparison between
the SDEs for WPPT and persistence plotted against look-ahead time for the period
June 2002 to May 2003. As can be seen, WPPT performs much better than persis-
tence, with the exception of the first few hours. The WPPT SDE rises rather slowly
with increasing look-ahead time in the range 5-10 per cent (of installed capacity).
The physical system is non-stationary, so the forecasting method used should be
able to adapt to changes in the physical system. Changes occur in the NWP models, the
number of wind farms or wind turbines within a wind farm that are not operating due
to forced or scheduled outages, extensions to wind farms, performance due to build-up
of dirt on turbine blades, seasonal effects, terrain roughness, and sea roughness for
offshore wind farms.
A wind power forecasting system which incorporates this adaptive property, in
that the model can fine-tune its parameters during online operation, is AWPPS
(ARMINES wind power prediction system), which was developed at ´ coles des
Mines in France. AWPPS integrates (i) short-term forecasts based on statistical
þ
Search WWH ::




Custom Search