Wind power resources in mountainous regions are conditioned on a vast variety
of factors influencing air flow. Complex topography causes various phenomena
such as localised thermal winds, acceleration due to tunneling and Foehn winds
interfering at a range of spatial scales and varying in time due to weather seasonality.
It increases the dimensionality of parameter space and adds additional
complexity to sampling strategies and monitoring network design for wind resource
assessment and location allocation for wind turbines. This work explores
an active learning approach to multiple kernel learning (MKL) to explore the highdimensional
space of topographic features influencing wind speeds. MKL allows
handling spatial heterogeneity and non-stationarity while providing physically interpretable
data-driven models useful for decision support. Our results on real data
from the Swiss Alps suggest the efficiency of MKL both for feature selection, predictive
modelling and sampling design, also showing that care has to be taken to
avoid over-fitting by over-localised terms in kernel dictionaries.
[1]
Greg Schohn,et al.
Less is More: Active Learning with Support Vector Machines
,
2000,
ICML.
[2]
Devis Tuia,et al.
Learning wind fields with multiple kernels
,
2011
.
[3]
N. Cristianini,et al.
On Kernel-Target Alignment
,
2001,
NIPS.
[4]
Nello Cristianini,et al.
A statistical framework for genomic data fusion
,
2004,
Bioinform..
[5]
Gunnar Rätsch,et al.
Large Scale Multiple Kernel Learning
,
2006,
J. Mach. Learn. Res..
[6]
David A. Cohn,et al.
Improving generalization with active learning
,
1994,
Machine Learning.