A novel insight into spatio-temporal variability of storm events for modelling hydrological processes at catchment scale based on machine learning

Hydrological processes are shaped by complex and distant processes characterised by high spatio-temporal variability. Being the first hydrological process, triggering the remaining ones, precipitation, or more precisely, storm events, have paramount importance on the subsequent evolution of the hydrological system. The spatio-temporal evolution of precipitation has received profound attention from scientists. This topic is commonly addressed in practical hydrological simulation by simple (pseudo) deterministic algorithms as form example Polygons of Thiessen or Krigging methods. In this work we present a novel approach based on two pillars: first by focusing on storm events instead of in aggregated precipitation values and second by spatially analysing the relationships among the recorded values aided by machine learning algorithms. With that aim we have retrieved precipitation records from 6 weather stations in Madrid city with hourly latency from January 2019 and 587 stations with 15 minutes latency from January 2004. We have extracted the observed storm events in any case and analysed the spatio-temporal patterns underlying the storm evolution thus observing the scarce representativity of the traditional methods being machine learning approaches better suited for providing representative data.  This work is part of the project TED2021-131520B-C21, supported by the MCIN/AEI/10.13039/501100011033 and the European U nion “NextGenerationEU”/PRTR.