Estimation statistique de donnees manquantes en inventaire du cycle de vie

The main objective of the research work is to improve the quality of life cycle inventory data by developing a method to estimate missing data and corresponding uncertainties. Contrary to process-based models of mass and energy balance, this approach consists of statistical estimators which model processes from relatively small samples of usually high variability. The research hypothesis is as follows: the so called kriging estimator allows the combined estimation of missing data and their uncertainties in such ways that are more reliable than other linear estimators. Borrowed from spatial statistics, kriging is an estimator with several advantages, the flexibility associated with a choice of model function and the exact estimator property. In other words, kriging shows no statistical errors when estimating observed values, no data is averaged out. An interpretation of the kriging parameters specific to the problems of data uncertainty, offers more advantages. One parameter of the covariance function accounts for small scale variations of the data and taken as a proxy for uncertainty. Whether it be the variety of data sources, the scarcity of the data itself or both, each and every source adds to data variability and uncertainty. The kriging system of equations is therefore modified such as to integrate a factor of uncertainty specific to each observations. Comparisons between the modified and conventional forms of kriging can be drawn. The procedure is based on the relationship between technical specifications, more readily available independent variables, and the dependent material and energy flows of the processes under consideration. Such material and energy requirements as well as emissions are estimated over the entire life cycle of products and processes. The needs for additional data are relatively low compared to other approaches, namely extended input output analysis.;For many products, processes and services, electricity generation and consumption account for a sizable share of the impacts. Hydroelectricity in particular is poorly represented within existing inventory data since production facilities vary considerably from one location to another. In other words, generic hydropower plants do not exist. Contrary to inventory flows, the technical specifications or characteristic variables of hydropower plants, such as the installed capacity, annual production or surface area of adjacent reservoirs, are usually publicly available. The kriging model is first tested on a data set which represents windmills of varying power capacity before it is applied to hydroelectricity. The experiment is divided according to data availability, on one hand the energy and materials required during construction, operation and maintenance of hydropower plants.;The results show that estimation of inventory data can be improved thanks to kriging. When comparing different forms of kriging and linear regression, the kriging estimates are not only more precise but the standard deviations also cover the data more accurately. Where the observed data are incomplete, that is where inventory flows are missing for part of the observations, the estimation errors are lower for kriging than linear regression. Moreover, univariate kriging of inventory flows based on two characteristic variables, shows lower errors than its multivariate kin, cokriging. On average the statistical errors calculated from cross-validation are lower for kriging than they are for linear regression, whether the observed data are complete or not. The application of several characteristic variables improves the quality of the estimates when they are positively correlated. In addition, the modified form of kriging which accounts for degrees of uncertainty specific to each observations, results in a reduction in the variations of the estimated inventory data. That is, data variability is incorporated directly in the model. Estimates closer to more reliable observations are shown to be less uncertain and vice versa. For each of the data sets, different relationships between dependent and independent variables are tested, for example the linear, exponential, spherical and cubic covariance functions as well as a range of parameter values.;For the analysis of electrical generation technologies, these results imply better estimates for data that are difficult to sample and therefore a simplified data collection process. In the case of site specific or variable processes such as hydroelectricity, the estimation of inventory data with kriging accounting for such data variability, proves more representative of the geographical or technological context. The quality of inventory data is consequently higher. Even if kriging has several advantages and its estimation errors are lower on average, some limitations to its application exist. (Abstract shortened by UMI.).