Optical turbulence forecasts at short time-scales using an autoregressive method at the Very Large Telescope

Abstract

In this study we demonstrate that we can provide forecasts of all the main astroclimatic parameters (seeing, wavefront coherence time, isoplanatic angle, and ground-layer fraction) on time-scales of 1 and 2 h (the most critical ones for the service mode) with a root-mean-square error (RMSE) that is smaller than or, at worst, comparable to the instrumental uncertainty (i.e. the standard deviation between instrument estimates). The seeing RMSE is 0.08 arcsec. Results are achieved thank to the use of the autoregressive method (AR) in our automatic forecast system and the study is applied to the Very Large Telescope (VLT). The AR method is a hybrid method taking into account forecasts of a non-hydrostatical mesoscale model jointly with real-time observations made in situ. We demonstrate that the AR method allows an improvement in forecast performance of roughly a factor of three or more with respect to the standard forecasts at a long time-scale (beginning of the afternoon for the coming night), depending on the parameter and the time-scale (1 and 2 h). The AR method also allows roughly a factor of two gain with respect to prediction by persistence. We also show that the AR method provides significantly better performance than a random-forest machine-learning algorithm. An extended analysis of the AR performance is provided following different strategies. Results achieved in this study are therefore very promising and tell us that we can provide real assistance to the service mode of the VLT instrumentation supported by adaptive optics systems.