Generative models represent a fascinating group of methods that can generate new samples (such as images) with similar properties to the data used to train the generative model. The models have also been used to perform generative forecasts, such as the next frames of a video or the weather for the next hours. However making these methods perform well in such cases is challenging. In this project, we will explore the potential of using simpler models to learn the temporal properties, and only use generative models to learn the spatial dependencies.
The temporal properties can for example be learned using time series models such as ARMA models. We will further expect complex spatial patterns in the forecasting errors from these models, and we suggest to learn these patterns using generative models. The suggested model can further be compared against generative models learned on the data from scratch.
Goal
Develop a new type of generative forecasting models.
Learning outcome
- Machine learning
- Generative models
- Time series modeling
- Real-world applications
Qualifications
- Python programming
- Knowledge about machine learning is an advantage
Supervisors
- Hugo Hammer
- Michael Riegler
References
- Leinonen, J., Hamann, U., Nerini, D., Germann, U., & Franch, G. (2023). Latent diffusion models for generative precipitation nowcasting with accurate uncertainty quantification. arXiv preprint arXiv:2304.12891.