Towards a Weather-Based Prediction Model For Starlink Throughput

Low Earth Orbit (LEO) Satellite Networks, such as SpaceX’s Starlink, are rapidly advancing technologies in the telecommunications sector. While promising global internet access, the link parameters, such as throughputs and latency, are instable and susceptible to weather conditions. In this paper, we aim to make the Starlink performance more predictable. Precisely, we compare the capabilities of different Machine Learning (ML) models to predict Starlink’s download and upload throughput based on weather conditions such as rain and clouds. We focus our analysis on feature importances to determine which weather factor is a good predictor. We found that the Random Forest (RF) model has the best predictive power for both the download (R(exp 2) = 0.47) and upload (R(exp 2) = 0.13) throughput. Furthermore, we identified rainfall as the most important predictive factor, followed by cloudiness. Our results indicate that the Starlink download throughput is more susceptible to weather conditions than the upload throughput, validating results of earlier studies. Generally, weather conditions alone are not sufficient to precisely predict Starlink throughputs, highlighting the complexity of impact factors on satellite links.