On the potential and limitations of Bayesian ensemble algorithms for the decomposition of time series generated by tokamak diagnostics

The vast majority of signals generated by tokamak diagnostics are in the form of time series. Consequently dealing with time-indexed data is a major task, to be tackled daily by both experimentalists and analysts. Decomposing a time series in terms of seasonal components, trends, change-points and noise is therefore a crucial activity, per se and as a preliminary step to further investigations. In the present work, the Bayesian ensemble approach to model decomposition of time series, originally developed for remote sensing of the earth, is applied to various global measurements routinely available in tokamak devices. Among the competitive advantages of the methodology, particularly relevant are its holistic view of the data and the independence from the details of the statistical algorithms and models. The potential of the technique, implemented by the BEAST code, has been assessed with both synthetic signals and experimental data. The approach proves to be very reliable in modelling trends and determining the time locations of abrupt changes even of strongly oscillatory components, such as ELMs and sawteeth. Deployment to assess small drifts confirms the lack of stationarity in tokamak high performance discharges. The difficulties of modelling the details of the sawteeth and irregular ELMs indicate the need to improve the method to deal with seasonal components of complex harmonic content and/or varying frequency. However, the available routines are already very effective in determining the times changes in the ELM regimes.