In recent years, with geopolitical conflicts increasing the cost of fossil fuels and the rapid growth of renewable energy production , electricity prices are experiencing extreme volatility. The complexity of forecasting electricity prices has grown, as often violent price movements are driven by the lack of or oversupply of renewable energies, and this in turn is solely driven by the suitability of weather conditions for production. Electricity price forecasting has been largely studied in both tabular machine learning approaches and in deep learning approaches, and largely with autoregressive approaches on price history itself or using coupled electricity market features, and sometimes with small sets of weather features, but there exists a gap at using weather data at a high spatial resolution as input, which is key in high renewable penetration markets like Germany where solar and wind production often meet electrical demand. This thesis aims to develop a deep learning framework to extract value from weather data from 10940 locations across Germany to forecast electricity prices. The dimensions of this kind of data demands careful model design: This thesis explores several variants of graph neural networks for spatial processing, Seq2Seq modules for temporal processing, and the effectiveness of their combination as a spatiotemporal hybrid deep learning architecture. The approach outlined in this thesis will be benchmarked against SOTA standards and deployed internal company models.
In recent years, with geopolitical conflicts increasing the cost of fossil fuels and the rapid growth of renewable energy production , electricity prices are experiencing extreme volatility. The complexity of forecasting electricity prices has grown, as often violent price movements are driven by the lack of or oversupply of renewable energies, and this in turn is solely driven by the suitability of weather conditions for production. Electricity price forecasting has been largely studied in both tabular machine learning approaches and in deep learning approaches, and largely with autoregressive approaches on price history itself or using coupled electricity market features, and sometimes with small sets of weather features, but there exists a gap at using weather data at a high spatial resolution as input, which is key in high renewable penetration markets like Germany where solar and wind production often meet electrical demand. This thesis aims to develop a deep learning framework to extract value from weather data from 10940 locations across Germany to forecast electricity prices. The dimensions of this kind of data demands careful model design: This thesis explores several variants of graph neural networks for spatial processing, Seq2Seq modules for temporal processing, and the effectiveness of their combination as a spatiotemporal hybrid deep learning architecture. The approach outlined in this thesis will be benchmarked against SOTA standards and deployed internal company models.
Horizon-stable multi-day day ahead electricity price forecasting
BRULE GALLEGUILLOS, DIEGO ALONSO
2025/2026
Abstract
In recent years, with geopolitical conflicts increasing the cost of fossil fuels and the rapid growth of renewable energy production , electricity prices are experiencing extreme volatility. The complexity of forecasting electricity prices has grown, as often violent price movements are driven by the lack of or oversupply of renewable energies, and this in turn is solely driven by the suitability of weather conditions for production. Electricity price forecasting has been largely studied in both tabular machine learning approaches and in deep learning approaches, and largely with autoregressive approaches on price history itself or using coupled electricity market features, and sometimes with small sets of weather features, but there exists a gap at using weather data at a high spatial resolution as input, which is key in high renewable penetration markets like Germany where solar and wind production often meet electrical demand. This thesis aims to develop a deep learning framework to extract value from weather data from 10940 locations across Germany to forecast electricity prices. The dimensions of this kind of data demands careful model design: This thesis explores several variants of graph neural networks for spatial processing, Seq2Seq modules for temporal processing, and the effectiveness of their combination as a spatiotemporal hybrid deep learning architecture. The approach outlined in this thesis will be benchmarked against SOTA standards and deployed internal company models.| File | Dimensione | Formato | |
|---|---|---|---|
|
BruleGalleguillos_DiegoAlonso.pdf
Accesso riservato
Dimensione
7.38 MB
Formato
Adobe PDF
|
7.38 MB | Adobe PDF |
The text of this website © Università degli studi di Padova. Full Text are published under a non-exclusive license. Metadata are under a CC0 License
https://hdl.handle.net/20.500.12608/115823