Περίληψη |
Environmental concerns have encouraged the adoption of renewable energy alternatives to reduce
greenhouse gas emissions across the world. As a consequence of that wind farms have been installed
in several locations across Greece for the production of wind power. Wind farm management and
control, power distribution planning, storage capacity management, and system's dependability, all
benefit from reliable wind speed forecasts.
In this thesis, the problem of the wind speed forecast has been approached through different
preprocessing techniques, that involve scaling, smoothening or data augmentation, as well as Artificial
Neural Network models that are based on Gated Recurrent Units (GRUs). Moreover, comparisons were
made for different Datasets with consisting of various sampling steps and Datasets, such as Jena
Climate, that contain significantly more observations.
The comparisons that were made for the preprocessing techniques indicated that the smoothening
approach managed to capture more accurately the dynamic structure of the data and perform robust
predictions, even for up to 10 – 12 hours ahead. At the same time, the different Neural Network
Architectures that were proposed had no significant differences with respect to their performance.
Comparison between datasets with different sampling steps (1 hour and 10 minutes) indicated no
systematic differences as a result of the stochastic nature of the timeseries. Transfer Learning strategy
turned out to behave similarly to the original models for short term predictions, where the long term
forecasts appeared to be more robust. Jena Climate dataset, may indicate a potential increase in
accuracy of forecasts with the condition that more observations will be added in the datasets at hand.
|