Development of 60-min to 1-min integration time conversion model and application of machine learning for time-series attenuation prediction in tropical location

The demand for locally measured rain rate data with 1-min integration time continues to grow due to its accuracy in estimating rain-induced attenuation. However, the scarcity of local observatory weather stations due to the acquisition, maintenance, and the high cost of measuring equipment have resulted in the adoption of satellite data in most locations. Satellite-borne radars provide wide coverage but have low resolutions with higher integration time, hence the need for conversion to lower integration time. In this work, ten years (2005–2014) of rain rate satellite data obtained from the Tropical Rainfall Measuring Mission (TRMM) have been converted from the default 60 min to 1 min of integration time using a newly developed conversion model. The research locations are Akure, Ikole-Ekiti, Ogbomosho, and Oshogbo, located in the southwestern part of Nigeria. The developed model was validated using ground-based measured two years (2009–2010) rain rate data of 1 min integration time. The performance of the developed 60-min to 1-min integration time model when compared with other models showed an improvement with a good R² of 0.92, a lower prediction error of 5.8 %, a root mean square error of 5.79 %, and a statistically significant smallest p-value of 0.005, indicating the strongest evidence against the null hypothesis. The converted 1-min rainfall rate data was applied to the Synthetic Storm Technique (SST) rain-attenuation model to generate time series rain-induced attenuation. The complimentary cumulative distribution function (CCDF) of the study areas depicted attenuation values greater than 10 dB at 0.01 percent of the time, which could be threatening to radio wave propagation. A further novel technique of predicting time series attenuation was applied by testing three forecasting models, namely SARIMA, ARIMA, and SVM, using machine learning. The SVM outperformed other models with a better R², MAE, RMSE, and confidence intervals of 0.89, 1.64 %, 2.14 %, and [2.4544, 5.9273], respectively. The results will be found applicable to mobile networks at 5G and 6G systems operating at high frequencies, as well as terrestrial communication links like microwave point-to-point.