Characterisation of large-scale urban fire emissions by inverse modelling

Large-scale fires such as warehouse fires that have occurred in recent years or dramatic accidents like the Paris Notre-Dame Cathedral fire in 2019 have stressed the need to develop means of assessing the toxicity risks to the population and the environment of smoke plumes. A key challenge is to quickly provide the authorities with information on the areas impacted by the plume and the pollutant concentration levels to which the population is likely to be or to have been exposed. The Laboratoire Central de la Préfecture de Police (LCPP) aims to deploy a number of devices for measuring pollutants and tracers of smoke combustion during a fire. Subsequently, the application of an atmospheric dispersion model within the framework of a data assimilation approach should provide a source characterisation and a finer estimate of the concentration levels at points of interest.

To characterise the source, noticeably the released mass of pollutants and the emission height linked to a plume rise, an inverse problem method has been implemented. It is based on a Bayesian Markov Chain Monte Carlo (MCMC) technique meant to quantify the uncertainties associated with the emission estimation. Since the emission height strongly influences the atmospheric dispersion in the vicinity of the source, two approaches are used to estimate it. The first one consists in finding the emission time rate for each considered height and the second one consists in focussing on a single emission height and its associated emission time rate using a discrete distribution to describe the vertical profile. We use the Lagrangian Parallel Micro Swift Spray (PMSS) model developed by AriaTechnologies fed with meteorological fields provided by Météo-France to represent the atmospheric dispersion of smoke.

Our inverse method is applied to a large warehouse fire that occurred in Aubervilliers near Paris in 2021 using real observations. Abnormal concentrations of particulate matter were recorded, with a peak at 160 µg.m^-3, located in the centre of Paris about 6 km from the source. They were collected by the LCPP and AirParif, the local air quality agency, and are used to retrieve the emission with a quantification of uncertainties and a sensitivity analysis of model error. The resulting emission height of the source, mainly between 200 and 300 m, coincides with the terrain observation for an emission rate of less than 1000 kg/h throughout the duration of the fire. A sensitivity analysis to the initial approximation of the source (the prior) shows its importance. It suggests to improve our method by incorporating the statistical parameters of the observation error into the MCMC method.