Prediction of Atmospheric Noise Temperature at the Deep Space Network with Machine Learning

Ka-band (32 GHz) communications links utilized by the National Aeronautics and Space Administration (NASA) flight missions for science downlink are susceptible to degradation due to weather. In this study, a customized real-time forecast system has been developed to predict zenith atmospheric noise temperature (Tatm) at the Deep Space Network (DSN) tracking sites using machine learning (ML). A random forest model is trained with the Global Forecast System (GFS) forecast and analysis datasets in addition to the Tatm measurements derived from on-site advanced water vapor radiometers (AWVR). The real-time forecast uncertainty is quantified for different error regimes using the Self-Organizing Map method. The results show that the Root Mean Square Error (RMSE) of the 24-hour Tatm prediction at Goldstone, CA increases with the increase of Tatm. At 90% of the time, the forecasts have RMSE (bias) of less than 3.50 K (0.16 K) for fair-weather conditions with Tatm < 16 K. In comparison to the current approach in designing Ka-band communications links, application of weather forecasts can increase data return to the downlink for 80% of the time. A downlink gain of up to 2.43 dB (75% more data) can be realized at 10 elevation angle when Tatm = 8.32 K.