Uncertainty estimation of air quality simulations at urban scale – ESTIMAIR
Uncertainties of air quality simulations at urban scale
Propagation of uncertainties in a dynamic transport affectation model at urban scale, in the computation of the associated pollutant emissions and in the atmospheric dispersion of the pollutants in a city<br />
Air quality simulations at urban scale depend on city geometry, meteorological conditions, background pollution (imported from distant locations) and urban emissions of atmospheric pollutants, largely due to road traffic. Based on these data, an urban air quality model can estimate pollutant concentrations at city scale, for each hour and down to street resolution. These simulations are subject to high uncertainties because of the data inaccuracies and the model shortcomings. It is important to quantify the resulting uncertainties in order to provide confidence levels to the simulations and, in particular, to compute the variances and covariances of the simulation errors which are key inputs to data assimilation schemes.<br />
A first step is the implementation of the full simulation chain to compute the concentrations of the atmospheric pollutants at urban scale. Three computation stages are required: the traffic assignment on the road network, the computation of the associated emissions and the dispersion of the atmospheric pollutants. In order to produce appropriate traffic simulations, the dynamic traffic affectation model LADTA is used. The emissions are computed following the COPERT approach. Finally, the urban air quality model SIRANE is used to simulate the dispersion of pollutants in the streets and over the city. These simulations show very high computational costs. However, uncertainty propagation can require the generation of a large ensembles of simulations in order to properly represent the uncertainties. The design of metamodels, for traffic assignment and atmospheric dispersion, allows one to greatly reduce the computational burden and therefore to generate large ensembles of simulations. The calibration of these ensembles should finally be carried out, using the observations for traffic (traffic counters) and for atmospheric pollutant concentrations (fixed air quality monitoring stations).
We obtained the required data to apply our work to the city of Clermont-Ferrand (France). A major part of our work was dedicated to the simulation of road traffic, using the dynamic model LADTA. At rush hour, the simulations are similar to those of the reference, static model PTV VISUM. In addition, the dynamic model reproduces well the temporal profile of traffic volumes which are observed at more than 500 traffic counters of the city of Clermont-Ferrand. Nevertheless, the spatial distribution of traffic is not accurately reproduced, and the results show large uncertainties (possibly above 50%). Additionally, the computation of the emissions is highly sensitive to the composition of the vehicle fleet. As for the atmospheric dispersion, a complete one-year air quality simulation for the city of Clermont-Ferrand has been carried out with the SIRANE model. A reliable air quality metamodel was generated for the full city of Clermont-Ferrand. This metamodel is able to reproduce the observations as well as the original model. In order to compute one map of pollutant concentrations, the original model requires several minutes while the metamodel takes only 50 milliseconds.
Further work will consist in assigning traffic and computing the associated emissions for a complete year. The resulting emissions will serve as inputs to air quality simulations. Following the objective of the uncertainties propagation in the complete simulation chain, a metamodel will be generated for the traffic assignment model LADTA. Its ouputs will be used by an air quality metamodel. It will therefore be possible to generate large ensembles of simulations (for traffic and then for air quality), and to study their calibration based on observational data.
A paper on the generation of an air quality metamodel has been submitted.
This project aims to quantify the uncertainties of the pollutant concentrations that are computed by an operational urban air quality model. The uncertainties refer to the range of values that the errors (i.e., the discrepancies between the model outputs and the true values) can take. These errors are usually modeled as a random vector, whose probability density function is the complete description of the uncertainties. Our strategy to approximate this probability density function is the generation of an ensemble of simulations that properly samples the errors.
The application is air quality simulation across Clermont-Ferrand, using a dynamic traffic model to compute traffic emissions and using an atmospheric chemistry-transport model that explicitly represents the streets of the city. Based on the emission data, meteorological conditions and background pollutant concentrations, the air quality model computes every hour the concentration fields (across the whole city) of several air pollutants, especially dioxide nitrogen and particulate matter. As a result of the complexity of atmospheric phenomena and the limited observations, the simulations can show high uncertainties which need to be estimated. Our objective is to propose a tractable approach to provide uncertainty estimations along with any urban simulation. The approach should apply to short-term forecasts as well as long-term simulations (e.g., for impact studies).
One major uncertainty source lies in the traffic emissions. We will carefully estimate the uncertainties of traffic assignments in the streets and of associated pollutant emissions. Using multiple simulations of a state-of-the-art dynamic traffic model, an ensemble of traffic assignments will be generated. The ensemble will be calibrated with traffic observations so that it should be representative of the uncertainties of the traffic model. The associated ensemble of pollutant emissions will provide inputs to the air quality model. An ensemble of air quality simulations will be generated, using the different traffic emissions, using perturbed input data (Monte~Carlo approach) and possibly a multimodel approach. This ensemble will also be calibrated using observations of pollutant concentrations in the air. The air quality model is a high-dimensional model with high computational cost. In order to generate an ensemble of simulations, it is necessary to reduce the computational costs. Consequently a part of the project deals with the reduction of the air quality model.
This project is proposed in a context of increasing use of numerical air quality models at urban scale. The models are used for daily forecasts, for assessment of long-term exposure of populations to pollution, for the evaluation of the impact of new regulations, ... We will propose methods that can be applied in an operational context to the core modeling chain, from traffic assignment to atmospheric dispersion. The scientific results of the project will be integrated in an operational modeling system that is currently used for many cities in France and abroad.
Monsieur Vivien MALLET (Institut national de recherche en informatique et en automatique)
The author of this summary is the project coordinator, who is responsible for the content of this summary. The ANR declines any responsibility as for its contents.
LVMT Laboratoire ville mobilité transport
LMFA Laboratoire de mécanique des fluides et d'acoustique
Inria Paris - Rocquencourt Institut national de recherche en informatique et en automatique
Help of the ANR 414,800 euros
Beginning and duration of the scientific project: August 2013 - 48 Months