Deep-learning-based prediction of precipitable water vapor in the Chajnantor area

Context. Astronomical observations at millimeter and submillimeter wavelengths strongly depend on the amount of precipitable water vapor (PWV) in the atmosphere, which directly affects the sky transparency and decreases the signal-to-noise ratio of the signals received by radio telescopes. Aims. Predictions of PWV at different forecasting horizons are crucial to supporting the telescope operations, engineering planning, scheduling and observational efficiency of radio observatories installed in the Chajnantor area in northern Chile. Methods. We developed and validated a long short-term memory (LSTM) deep-learning-based model to predict PWV at forecasting horizons of 12, 24, 36, and 48 hours using historical data from two 183 GHz radiometers and a weather station in the Chajnantor area. Results. We find that the LSTM method is able to predict PWV in the 12- and 24-hour forecasting horizons with a mean absolute percentage error of ~22% compared to ~36% for the traditional Global Forecast System method used by the Atacama Pathfinder Experiment, and the root mean square error is reduced by ~50%. Conclusions. We present a first application of deep learning techniques for preliminary predictions of PWV in the Chajnantor area. The prediction performance shows significant improvements compared to traditional methods in 12- and 24-hour time windows. We also propose strategies to improve our method on shorter (<12 hour) and longer (>36 hour) forecasting timescales.