Research on wind power scenario generation method based on dynamic spatiotemporal graph generative adversarial network
Feng, Yuwei (2026)
Kandidaatintyö
2026
School of Energy Systems, Sähkötekniikka
Kaikki oikeudet pidätetään.
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi-fe2026061166597
https://urn.fi/URN:NBN:fi-fe2026061166597
Tiivistelmä
The widespread application of wind energy within the power grid has increased the demand for power system security and stability in the presence of uncertainties and volatility within the wind energy supply. The power generated from wind energy depends largely on environmental factors, such as wind velocity, wind direction, and changes in weather conditions within the region. As such, the uncertainty inherent in wind power cannot be effectively captured using a single deterministic forecasting model. For that reason, a scenario generation for wind power is required.
Previous techniques used in the scenario generation process included statistical models and some forms of deep learning algorithms that could capture uncertainty in wind power. However, these models may lack the capacity to capture temporal dynamics, spatial correlation, and evolving dependencies between the multiple wind farms. To solve this challenge, this thesis aims at developing a wind power scenario generation technique using DSTG-GAN.
In the approach, wind farms are taken as graph nodes, and correlations among multiple wind farms are modelled via graph edges. By combining graph-based modelling for spatial aspects and features extracted through time, the generator learns the spatiotemporal characteristics of multi-wind farm power data, whereas the discriminator evaluates consistency in distributions between the generated and real scenarios. Furthermore, physical and statistical constraints are incorporated in the process to increase the quality of the generated data.
As seen from experiments, the DSTG-GAN framework can produce realistic wind power scenarios based on temporal trends, statistical properties, and spatial correlations observed in real data. Even though there are minor deviations, such as a bit of underestimation in areas where high wind power occurs and sometimes local variations, the model proves the viability of the idea of dynamic spatiotemporal graph modelling for generating wind power scenarios.
Previous techniques used in the scenario generation process included statistical models and some forms of deep learning algorithms that could capture uncertainty in wind power. However, these models may lack the capacity to capture temporal dynamics, spatial correlation, and evolving dependencies between the multiple wind farms. To solve this challenge, this thesis aims at developing a wind power scenario generation technique using DSTG-GAN.
In the approach, wind farms are taken as graph nodes, and correlations among multiple wind farms are modelled via graph edges. By combining graph-based modelling for spatial aspects and features extracted through time, the generator learns the spatiotemporal characteristics of multi-wind farm power data, whereas the discriminator evaluates consistency in distributions between the generated and real scenarios. Furthermore, physical and statistical constraints are incorporated in the process to increase the quality of the generated data.
As seen from experiments, the DSTG-GAN framework can produce realistic wind power scenarios based on temporal trends, statistical properties, and spatial correlations observed in real data. Even though there are minor deviations, such as a bit of underestimation in areas where high wind power occurs and sometimes local variations, the model proves the viability of the idea of dynamic spatiotemporal graph modelling for generating wind power scenarios.