Xgboost Based Regression Forecast for ACM on Q/V-band Satellite Links

This paper presents a comparison between classical adaptive coding and modulation (ACM) with fixed modulation and coding (ModCod) margins and two machine learning (ML)-based approaches: univariate and multivariate forecasting models. Both ML approaches are based on variants of Extreme Gradient Boosting (XGBoost) to predict signal-tonoise ratio (SNR) time series, aiding ACM switching decisions. The evaluation of the ACM algorithms is conducted using two years of Q/V-band channel data recorded at the ground station in Graz, Austria, using the Alphasat TDP5 Aldo Paraboni Q/Vband payload. The results demonstrate that the multivariate forecasting model outperforms both the classical ACM algorithm and the univariate forecasting model in terms of spectral efficiency. Additionally, the multivariate model eliminates the need for direct SNR estimation by using easily measurable parameters from the modem.