Multi-Horizon Probabilistic Forecasting of Rain-Induced Signal Degradation for Weather-Aware Radio Telescope Operations in a Tropical Climate

Rain fade is the dominant weather disturbance for high-frequency radio telescopes in the tropics, where convective storms develop within minutes and long-term statistical models such as ITU-R P.618 offer no short-term guidance. We present the first site-specific multi-horizon forecast of rain-induced signal degradation for a radio telescope in a tropical climate, built from 65 days (1249 h) of 30-s signal-to-noise ratio (SNR) telemetry from the KMITL 12-m Ku-band antenna in Bangkok, co-located with a 16-s surface weather station. An attention-based network with quantile output and a gradient-boosting ensemble forecast the apparent attenuation (the SNR loss from clear sky; a radiometric correction for rain-emitted noise yields the path attenuation) 5–60 min ahead under day-blocked cross-validation. The point skill against persistence reaches +0.13 at 60 min but is near zero at 5–15 min under rain: the slant-path signal itself leads every surface predictor, including the rain gauge by about 5 min. The forecast value appears instead at the event level. At matched false-alarm rates, quantile-based warnings detect 11 to 13 of 13 fade events with a 20–60 min median lead time, against 9 of 13 and under 20 min for reactive monitoring. A scheduling replay quantifies when forecast-based deferral beats reactive operation, and the fade climatology, lead–lag physics, and 17.7-dB (apparent) case study characterize the site itself.