Probabilistic Wind Power Forecasting with Deep Neural Sequence Models
Permanent lenke
https://hdl.handle.net/10037/30228Dato
2023-05-27Type
Master thesisMastergradsoppgave
Forfatter
Svenøe, SofieSammendrag
As the world strives to fulfill the goal of zero-emission established during the Paris agreement, an increasing amount of wind power is integrated into the liberalized electricity markets. With this escalation comes the need for wind power forecasting (WPF) due to the intermittent nature of wind, and WPF is therefore becoming an important field of study to successfully incorporate wind power to the electricity market. Given the rapid growth of machine learning, deep learning and probabilistic forecasting has emerged as good alternatives for WPF due to their non-linear processing methods and their ability to model uncertainties.
In this study, two probabilistic deep learning networks and a statistical model are tested as WPF models for a 54 MW wind power park. The models are trained to predict for the day-ahead and intraday electricity market, which respectively has 12-26~h and 1-24~h as associated forecasting horizons. Historical wind power production and Numerical weather predictions (NWP) are used as input to the WPF models. NWPs are modeled from the MEPS model, operated by the Norwegian Meteorology Institute (MET Norway).
The tests show that the two neural network models Temporal Fusion Transformer, and DeepAR, produces better predictions than the statistical model, SARIMAX, for the day-ahead market. The neural networks achieved P50/P90-Risk respectively of 0.153/0.081, and 0.175/0.091. While, for the intraday market, the models DeepAR, and SARIMAX performed substantially better than the Temporal Fusion Transformer, with P50/P90-Risk of respectively, 0.111/0.056, and 0.184/0.099. This implies that Transformer sequence models perform best on long-term forecasting, whereas autoregressive models still perform best on short-term forecasting.
Forlag
UiT The Arctic University of NorwayUiT Norges arktiske universitet
Metadata
Vis full innførselSamlinger
Copyright 2023 The Author(s)
Følgende lisensfil er knyttet til denne innførselen: