Pronóstico de lluvia en una cuenca de alta montaña basado en técnicas de aprendizaje automático (Machine Learning) y datos del radar meteorológico de medición de lluvias de banda X CAXX

Abstract

Literature related to the rainfall forecast in highlands using Machine Learning techniques is not extensive, especially in tropical Andean regions such as Ecuador. One of the possible reasons is the limited data due to the difficulty of implementing hydro-meteorological stations in these zones. This also because of the complex topography, which at the same time makes it difficult to obtain acceptable forecasting results since there is a high level of temporal rainfall variability. In order to obtain better rain observations in Ecuador, a network of meteorological radars called RadarNet-Sur was implemented; each radar provides observations of reflectivity which can be used to derivate rain rates. One of the most essential zones that one of the radars covers is the city of Cuenca (> 600.000 inhabitants) where extreme events such as flooding’s and river overflows can have catastrophic consequences. Therefore, the development of rain forecasting models, in order to prevent the effects of the extreme events previously mentioned, is crucial for the decision-makers. In this thesis, Support Vector Machine and Random Forest techniques were used to forecast one-hour reflectivity by using the data of Radar CAAX and converting it into rain rate. The results indicate that it is possible to learn from the data only by using a reflectivity variable captured by the X-band radar. On the one hand, the random forest models provide values approximated to the optimum for the metrics PCC, BIAS and RMSE, which allow forecasting reflectivity in order to transform it into rain rate.