Year: 2024
Type: dissertation
Abstract: Increased integration of photo-voltaic capacities in Europe has caused the electrical grid to become more difficult to maintain. In order to operate the grid in an efficient manner, its operators need to rely on accurate forecasts of the solar power generation. For this purpose, our work focused on predicting 2-hour ahead solar power generation at 15-minute intervals, a typical resolution requirement for solar power plants and system operation measurements. The modelling methodology in our work revolved around a state-of-the-art deep learning architecture called textit{Temporal Fusion Transformer} cite{tft} proposed by researchers at Google. The input to the model consisted of three different data sources: satellite imagery from the European Organisation for the Exploitation of Meteorological Satellites, numerical weather predictions from the European Centre for Medium--Range Weather Forecasts and solar power data from the Slovenian Transmission Network Operator ELES. In contrast, only one or two of these input sources are usually used in currently available research cite{comparison2022, intraday2021, history2018}. We made solar power generation forecasts for 233 different locations across Slovenia -- more extensive research regarding the number of photo-voltaic locations than what can be found in the literature on this topic cite{tft2022}. Temporal Fusion Transformer outperformed well-established benchmarks in solar forecasting by significant margins across all metrics and training settings considered. Our ablation studies also showed that the Temporal Fusion Transformer is capable of extracting a sufficient amount of information from satellite images meaning that such images can be used as a decent replacement for solar power data when it is not available.
Author: Stefanov, Dimitar
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7002363788
Cites:
Cited by:
Related to: