TSPRocket: A Fast and Efficient Method for Predicting Astronomical Seeing

Cheng-Qin Zhang, Weijian Ni

Abstract: When using large optical telescopes, astronomical seeing prediction is an essential and challenging work. It can provide observation tips for optical telescopes and help astronomical observers to arrange observation tasks reasonably. The traditional method of seeing prediction mainly uses meteorological models to capture atmospheric turbulence patterns and then implement seeing prediction through data analysis. In recent years, data-driven methods have made good progress in seeing prediction research. This paper proposes a new data-driven seeing prediction method called TSPRocket (Time Series Prediction RandOm Convolutional KErnel Transform), which uses a large number of convolution kernels to transform time series data, and then the transformed features are used to train a simple linear predictor. Because there is no need to calculate gradient when performing convolution calculation, TSPRocket is much more efficient than existing methods. In order to facilitate the research, we collected the observation data of a large astronomical telescope, and then carried out necessary data preprocessing. The experimental results show that our TSPRocket has higher accuracy and speed than baseline models.