Deep learning architectures in energy time series forecasting

OData support
Supervisor:
Gáspár Csaba
Department of Telecommunications and Media Informatics

Energy demand/load forecasting is a crucial topic for electricity providers because their ability to produce the energy, exceeds the costs of storing it. In advance to waste less, they need as precise estimations as possible.

This type of forecast is becoming more and more popular with the development of machine learning. With the usage of big training data, these algorithms are able to learn non-linears patterns. In the last decade, computing capacity is growing faster and faster, especially in the GPU technology, which provides a good opportunity to use more complex learning structures. The development of deep learning nets also becoming faster and faster each year, companies like Google, Microsoft, or Facebook are the leadership of the development process.

Deep learning is a set of algorithms designed to train artificial neural networks. It also means different net structures, and several techniques to denoise these nets, and prevent overlearning. During my work this time, I examined the capabilities of these structures, in the context of energetical time series’. I became familiar with the basics of the topic, and then constructed models for short- and mid-time forecasting, using various datasets.

Downloads

Please sign in to download the files of this thesis.