Can Precipitation-Related Incidents on the French Railway Network be Predicted in Advance? : Internship Project Report

To improve flood mitigation, the French Railways (SNCF-Réseau) would like to use short-term forecasts to predict when a precipitation-induced incident would occur on the railway network. This study explored whether precipitation-induced incidents on the SNCF network between 2018-2022 can be predicted in advance. Precipitation accumulations from a gridded reanalysis dataset were matched with a log of incidents recorded by SNCF-Réseau. Local accumulation thresholds of a 10-year or 100-year event were tested to predict incidents. 24- and 72-hr precipitation accumulations were found to be weak – but not insignificant – predictors of the incidents provided (HSS<0.16). Filtering results based on antecedent moisture, using SPI-1, did little to improve results. Two case studies were then conducted to explore the cause of incidents, given various preceding rainfall amounts. In the Villaine case study, incidents occurred despite normal precipitation accumulations and antecedent moisture (SPI<1). Photos of incidents indicate that they have likely been caused by short, localised precipitation events, some of which may not have registered in the precipitation data used. In the Roya case, all incidents were registered after one extreme precipitation event. A hindcasting exercise - looking at past precipitation forecasts leading up to the event - showed that this event could have been forecasted with a lead time between 5 hours and 4 days, depending on the probability threshold used. Results indicate that many of the incidents were triggered by very intense and localised precipitation events. These events were underestimated, or not captured, by the precipitation products we used. Further studies would also benefit from more contextual insights on past incidents, in order to better categorise each incident according to their causation pathway and hydro-meteorological predictors. To this end, we been recommend to explore predictors of past incidents and use high-resolution precipitation observations and nowcasts. Also, recommendations have been made in further developing the risk management protocol to better gain insights on the hydro-meteorological context of each incident.