Summary:
The scope of the present doctoral dissertation is the spatio-temporal modeling
of meteorological parameters with the use of Artificial Neural Network (ANN)
models at Chania plain in Crete, an area characterized by complex topography.
The study is focused on the identification of the optimum type and architecture
of the developed ANN models in order to initially examine the relationship of
the synoptic scale circulation with local meteorological conditions and
consequently the spatial estimation and prediction of wind speed and ambient
temperature fields. Moreover, ANN models are developed for the spatio-temporal
statistical downscaling of ambient temperature.
The developed original atmospheric fields clustering and classification
methodology is based on competitive learning and the use of Self-Organizing
Maps (SOM). The methodology is applied to the fields of surface pressure (MSLP)
and 500hPa geopotential height (GH500hPa) for a synoptic scale grid, as well as
to the horizontal wind speed components at 10m and 850hPa levels (U10m, V10m,
U850hPaκαι V850hPa), surface temperature (T2m), dew point temperature at 2m
(Td2m) and 700hPa specific humidity (SH700hPa) for a grid centered over Greece.
The grid data are obtained from the ERA-Interim database with a 0,75°0,75°
spatial resolution. In total, 32 atmospheric patterns are identified, allocated
in an 84 map, and consequently, each pattern is related with the surface
meteorological conditions. An important advantage of this approach is that
neighboring centroids are interconnected and their relative position in the SOM
map is associated with specific features, such as seasonality, location the of
the pressure systems, the pressure gradient and their one-day conditional
transition probability, enabling the extraction of valuable information
regarding the evolution of atmospheric circulation.
The comparative study for the investigation of the ANNs ability to spatially
estimate wind speed and temperature is performed using experimental data from
six meteorological stations. The predictive accuracy of the feed-forward ANNs
with hyperbolic tangent sigmoid transfer functions (FFNN) is found to be
statistically significantly superior from the traditional linear spatial
interpolation methodologies and radial-basis function ANNs (RBF). In the case
of wind speed, the study demonstrates the significant effect of wind direction
along with the importance of the degree of representativity of
reference-station anemological data. The FFNNs are not used as ‘black-box’
models and they are proven to incorporate the spatial wind speed variability
and the effect of topography in the case of ambient temperature. In the case of
wind speed, the FFNN models are suggested to be used as a Measure – Correlate –
Predict methodology for wind energy applications.
As far as the prediction of the studied parameters is concerned, hourly data
from a single station are used along with theNonlinear Autoregressive Networks
(NAR). The comparative study is performed for different feedbacks (6, 12 and 24
hours) and highlights the importance of a high feedback degree, which leads to
an increase in forecast horizon. The models in the case of wind speed can be
used for short-term forecasting, while in the case of temperature for
medium-term ones. Furthermore, the forecasting error depicts a distinguished
seasonality pattern and is correlated with specific cyclonic patterns, where
the low-pressure center is located in northern Europe or the Mediterranean.
The ANN temperature spatio-temporal downscaling ensemble models consist of six
individual FFNNs, one for each forecast horizon. The study examines their
ability to spatiotemporally downscale ambient temperature at a single site,
from the six-hourly ERA-Interim gridded data. It is found that the predictive
accuracy of the model that incorporates a wide range of predictor variables
(T2m, T850hPa, U10m, U850hPa, V10m, V850hPa, SH1000hPa, SH700hPa, GH500hPa,
MSLP and Td2m) is superior compared to the single predictor (T2m) model. The
most important predictor variables are T2m, U10m, V10m, SH1000hPa and MSLP and
therefore they are related to the surface meteorological conditions. The
forecasting error is also characterized by seasonality and its higher values
are related with cyclonic atmospheric patterns.
The dissertation highlights the significance and the increased predictive
ability of ANNs in applied climatology, meteorology and wind energy
applications. Finally, the crucial factor for their optimum performance is the
appropriate selection of the ANN type and the corresponding architecture.
Keywords:
Artificial Neural Networks, Wind Speed, Temperature, Atmospheric Circulation, Complex Topography
