Estimation de l'incertitude sur les flux d'inventaire du cycle de vie - modélisation et développement de facteurs empiriques pour l'approche pedigree

Life cycle assessment (LCA) is a decision support tool that allows evaluating the potential environmental impacts of a product, service or system from the perspective of its whole life cycle: from resource extraction to end-of-life. LCA aims to compare the environmental performances of different systems or of the different life cycle stages of a same product. This knowledge allows legislators, the private sector or consumers to consider environmental aspects in their decisions regarding the assessed systems. As a decision support tool that relies on modelling, uncertainties are an intrinsic part of LCA models. The uncertainty of models and their inputs can be evaluated during a LCA. This evaluation can help improve the confidence in the results and in the decision the LCA results support. The most documented typology of uncertainty is uncertainty on life cycle inventory (LCI) parameters (LCI is the step of the LCA that aims to quantify all the exchanges between the system and the environment). These parameters are both uncertain – we don't know their “true” value – and variable – their value changes according to time, space and type of individual. There is presently no universal method to address LCI parameter uncertainty and variability. Only the ecoinvent LCI database offers a semi-quantitative approach to manage and quantify uncertainty on all the background data modelled in the database. This approach, called “the pedigree approach”, is based on the use of statistics and a pedigree matrix to combine two typologies of uncertainty. This approach quantifies 1) the basic uncertainty that represents the stochastic error ;and 2) the additional uncertainty that is due to the use of an imperfect data, imperfect according to the context study and according to how the datum was obtained. This pedigree approach, as applied in ecoinvent v2 has several limitations: the approach can only be applied if a datum with its uncertainty is modelled by a lognormal distribution, the uncertainty factors for both the basic and the additional uncertainties rely on experts' judgments rather than on empirical basis and the additional factors are unique whatever industrial sector. This thesis aims to address these limits by developing and legitimating, on a scientific basis, the semi-quantitative uncertainty management tool on uncertainty parameters that is the pedigree approach. In order to meet this global objective, three specific objectives are addressed: 1) legitimate the use of the lognormal distribution as a default distribution to model a parameter with its uncertainty; 2) determine how the pedigree approach can be applied to distributions other than lognormal; and 3) develop a data assessment structure that permits to determine empirically-based uncertainty factors and that, by activity sectors. By meeting these specific goals, this thesis makes several contributions to the modelling of uncertainty of LCI data. First of all, the analysis of the effects of the choice of default distributionto represent uncertainty of background LCI data shows that, while the choice can significantly affect the LCA results when product systems are assessed individually, it has no material effect on products system comparisons. Giving these results, the lognormal distribution can be used as a default distribution to model a datum and its uncertainty and that, if no information to specifically define the datum's distribution is available. The development of equations that combine basic uncertainty and additional uncertainty for all distributions in the ecoinvent database to model a datum and its uncertainty allows the practitioner and the data provider to choose the distribution that best fits the assessed datum. Lastly, with the use of Bayesian inference, a methodology that aims to refine uncertainty factors is developed. Applying this methodology allows to refine uncertainty factors whenever new data are accessible. This methodology is applied in the context of this thesis,which led to the development of empirically based, refined and sector-specific uncertainty factors. These factors are mostly higher than the ones presently used in ecoinvent. The developments related to the achievements of the goals are based on some assumptions that need to be considered when the results are used. Despite these assumptions, this thesis brings a complete justification of the use of the pedigree approach. This work is of interest only if the results are used by the practioners or the data developers who need them. Discussions are engaged with the ecoinvent centre to integrate the developed uncertainty factors into the database. The approach could also easily be used with other databases. Finally, this thesis contributes to the improvement of uncertainty modeling and to the knowledge on uncertainty on LCI parameters. This knowledge is necessary to systematically consider uncertainty in LCA and to consider uncertainty on results as a source of information on these results rather than a source of weakness.