Short-term rainfall prediction using MLA based on commercial microwave links of mobile telecommunication networks

Rainfall prediction is a major problem with considerable socio-economic, industrial, and environmental impacts. The expansion of mobile telecommunication networks around the world is being used as an alternative to the declining number of rain gauges. Many researches, using those networks, have been carried out to solve water related issues in particular, and to propose hydro-meteorological applications in general. However, the possibility to use mobile telecommunications networks for rainfall prediction is still at its premises. Machine learning algorithms and techniques have been widely proven to be effective for rainfall prediction, using different geo-physical and environmental variables. In this paper, we propose to use machine learning algorithms, namely ensemble methods, to predict rainy events and their corresponding rainfall depths based on signal levels attenuations along microwave links of commercial mobile telecommunication networks. A sample of four microwave links, extracted from a dataset containing commercial microwave links data from the Netherlands, is considered. This dataset contains minimum and maximum powers received by base transceiver stations over 15-min intervals, i.e. four records per hour. A radar rainfall dataset with a spatial resolution of 1 k m , and a temporal resolution of 5 min, is used as rainfall observations. The predictions are done at two levels. First, the nature (wet or dry) of upcoming 15-min periods is predicted. Second, rainfall depths are estimated for upcoming 15-min wet periods. The results obtained show a prediction accuracy between 72% and 93% for the prediction of upcoming periods with a prediction horizon between 1 and 60 min. The correlation coefficient between predictions of rainfall depths and radar rainfall observations is between 0.70 and 0.98, and the coefficient of determination between 0.72 and 0.90. In addition, the prediction horizon can be extended up to 5 h with a prediction accuracy above 60%. These results reveal the potential of microwave links of mobile telecommunication for short-term warning systems in general, and flood prediction in particular, as our models tend to be very accurate for the prediction of heavy rainy events.