This thesis focuses on exploring Denoising Diffusion Probabilistic Models (DDPM) in respect to time series data distributions. Knowledge about image generating DDPMs is used to create Time Series DDPMs that are capable of synthesizing time series data. Different statistical, visual metrics as well as an applicative test are introduced to evaluate the quality of the generative models. In conducting this applicative test, not only is the quality of the models further measured, but also the investigation is carried out to determine whether a trained DDPM can perform in a continual learning environment.
Taking DDPM and training them to synthesize time series data.
