Application and Analysis of Generative AI Algorithms for Probabilistic Energy Forecasts

Probabilistic forecasting is essential for the optimal operation of modern energy systems, particularly in the presence of volatile renewable energy sources. This thesis investigates the application of Generative Adversarial Networks (GANs) for probabilistic forecasting of day-ahead electricity prices. Three different Conditional Time Series GAN (CTSGAN) models are implemented, incorporating exogenous factors such as electricity load and renewable generation forecasts. The model’s performances are evaluated against traditional forecasting methods, including SARIMAX and LSTMs, using metrics like the Continuous Ranked Probabilty Score (CRPS) as well as a simple optimisation approach for a trading strategy, highlighting the advantages of GAN-based approaches in capturing complex dependencies and generating realistic probabilistic forecasts. The results demonstrate that CTSGANs outperform baseline models in terms of forecast sharpness and reliability as well as optimisation results, providing valuable insights for energy market participants and grid operators.