Περίληψη |
If the COVID-19 pandemic is any indication is that science’s unique approach can provide solutions to the most dire situations. That approach is much needed when it comes
to sustainability matters. Planet sustainability is shaping up to be ”the” problem of the
21st century and it has never been more prevalent among the scientific community due
to its high complexity. Sustainability is a social, economic, political and environmental
problem. One of the focal issues surrounding a sustainable ecosystem is its energy dependence. From production to consumption and from storage to distribution, energy is
a resource solely dependent on how it’s utilized by us. Renewable Energy Sources (RES)
have shown the potential to relieve energy demands by providing, clean, green energy but
due to the nature of renewable energy, the energy output from such sources is inconsistent
and an active topic among the scientific community. Now casting weather phenomena and
forecasting long term climate variability and essential variables like surface land temperature, has been well documented. Layers of complexity are added when the interest of
research is towards providing knowledge over temporal periods of months.
To further facilitate this research we used deep neural networks (DNN). Deep learning
architectures have revolutionized many scientific domains due to their superior performance and flexibility by being able to extract aggregated knowledge when enough training
data is available. In this thesis we utilized the state of the art Convolutional Long - Short
Term Memory (ConvLSTM) architecture in order to forecast land - atmosphere boundary temperatures at the region of Crete over the span of months. We explored three different bands, temperature of air at 2 meters, skin temperature and soil temperatures
between 0 and 7 cm below surface level provided by ERA-5 Land, which is a climate
reanalysis dataset produced by the European Centre for Medium-Range Weather Forecasts (ECMWF) ERA-5 providing a consistent view of the evolution of land variables
over several decades. We explored the different cases surrounding the time horizon of the
forecast.
In the first case we explored our models performance when we predicted the afore-
mentioned bands a month ahead. Such a prediction gathers not only scientific and environmental interest but interest from the general populous too. It can provide invaluable
information about energy consumption and grid on - off grid stability.
In the second case we look to further our research by providing knowledge for longer
horizons. We present results that account for 6 months ahead. By doing so we provide
information that can be used for RES resources allocation betterment of policy making
and more sustainable and efficient energy consumption if the need arise.
In the third case we explored how subtracting data after the training and fit routine can
affect future predictions in the span of one, two and three months. This study revealed
how a limitation on previous knowledge can significantly affect future observations.
|