Petroleum engineers generate reservoir simulation models to optimize production and maximize recovery. History matching is one of the methods used to calibrate the reservoir models. During traditional history matching, individual model variable parameters (permeability, relative permeability, initial water saturation, etc) are adjusted until the production history is matched using the updated reservoir model. However, this method of utilizing only one model does not help capture the full range of system uncertainty. Another drawback is that the entire model has to be matched from the initial time when matching for new observation data. Ensemble Kalman Filter (EnKF) is a data assimilation technique that has gained increasing interest in the application of petroleum history matching in recent years. The basic methodology of the EnKF consists of the forecast step and the update step. This data assimilation method utilizes a collection of state vectors, known as an ensemble, which are simulated forward in time. In other words, each ensemble member represents a reservoir model (realization). Subsequently, during the update step, the sample covariance is computed from the ensemble, while the collection of state vectors is updated using the formulations which involve this updated sample covariance. When a small ensemble size is used for a large, field-scale model, poor estimate of the covariance matrix could occur (Anderson and Anderson 1999; Devegowda and Arroyo 2006). To mitigate such problem, various covariance conditioning schemes have been proposed to improve the performance of EnKF, without the use of large ensemble sizes that require enormous computational resources. In this study, we implemented EnKF coupled with these various covariance localization schemes: Distance-based, Streamline trajectory-based, and Streamline sensitivity-based localization and Hierarchical EnKF on a synthetic reservoir field case study. We will describe the methodology of each of the covariance localization schemes with their characteristics and limitations.

The goal of any data assimilation or history matching algorithm is to enable better reservoir management decisions through the construction of reliable reservoir performance models and the assessment of the underlying uncertainties. A considerable body of research work and enhanced computational capabilities have led to an increased application of robust and efficient history matching algorithms to condition reservoir models to dynamic data. Moreover, there has been a shift towards generating multiple plausible reservoir models in recognition of the significance of the associated uncertainties. This provides for uncertainty analysis in reservoir performance forecasts, enabling better management decisions for reservoir development. Additionally, the increased deployment of permanent well sensors and downhole monitors has led to an increasing interest in maintaining 'live' models that are current and consistent with historical observations. One such data assimilation approach that has gained popularity in the recent past is the Ensemble Kalman Filter (EnKF) (Evensen 2003). It is a Monte Carlo approach to generate a suite of plausible subsurface models conditioned to previously obtained measurements. One advantage of the EnKF is its ability to integrate different types of data at different scales thereby allowing for a framework where all available dynamic data is simultaneously or sequentially utilized to improve estimates of the reservoir model parameters. Of particular interest is the use of partitioning tracer data to infer the location and distribution of target un-swept oil. Due to the difficulty in differentiating the relative effects of spatial variations in fractional flow and fluid saturations and partitioning coefficients on the tracer response, interpretation of partitioning tracer responses is particularly challenging in the presence of mobile oil saturations. The purpose of this research is to improve the performance of the EnKF in parameter estimation for reservoir characterization studies without the use of a large ensemble size so as to keep the algorithm efficient and computationally inexpensive for large, field-scale models. To achieve this, we propose the use of streamline-derived information to mitigate problems associated with the use of the EnKF with small sample sizes and non-linear dynamics in non-Gaussian settings. Following this, we present the application of the EnKF for interpretation of partitioning tracer tests specifically to obtain improved estimates of the spatial distribution of target oil.

L'assimilation de données vise à fournir aux modèles de prévision numérique du temps un état initial de l'atmosphère le plus précis possible. Pour cela, elle utilise deux sources d'information principales : des observations et une prévision récente appelée "ébauche", toutes deux entachées d'erreurs. La distribution de ces erreurs permet d'attribuer un poids relatif à chaque source d'information, selon la confiance que l'on peut lui accorder, d'où l'importance de pouvoir estimer précisément les covariances de l'erreur d'ébauche. Les méthodes de type Monte-Carlo, qui échantillonnent ces covariances à partir d'un ensemble de prévisions perturbées, sont considérées comme les plus efficaces à l'heure actuelle. Cependant, leur coût de calcul considérable limite de facto la taille de l'ensemble. Les covariances ainsi estimées sont donc contaminées par un bruit d'échantillonnage, qu'il est nécessaire de filtrer avant toute utilisation. Cette thèse propose des méthodes de filtrage du bruit d'échantillonnage dans les covariances d'erreur d'ébauche pour le modèle à échelle convective AROME de Météo-France. Le premier objectif a consisté à documenter la structure des covariances d'erreur d'ébauche pour le modèle AROME. Une assimilation d'ensemble de grande taille a permis de caractériser la nature fortement hétérogène et anisotrope de ces covariances, liée au relief, à la densité des observations assimilées, à l'influence du modèle coupleur, ainsi qu'à la dynamique atmosphérique. En comparant les covariances estimées par deux ensembles indépendants de tailles très différentes, le bruit d'échantillonnage a pu être décrit et quantifié. Pour réduire ce bruit d'échantillonnage, deux méthodes ont été développées historiquement, de façon distincte : le filtrage spatial des variances et la localisation des covariances. On montre dans cette thèse que ces méthodes peuvent être comprises comme deux applications directes du filtrage linéaire des covariances. L'existence de critères d'optimalité spécifiques au filtrage linéaire de covariances est démontrée dans une seconde partie du travail. Ces critères présentent l'avantage de n'impliquer que des grandeurs pouvant être estimées de façon robuste à partir de l'ensemble. Ils restent très généraux et l'hypothèse d'ergodicité nécessaire à leur estimation n'est requise qu'en dernière étape. Ils permettent de proposer des algorithmes objectifs de filtrage des variances et pour la localisation des covariances. Après un premier test concluant dans un cadre idéalisé, ces nouvelles méthodes ont ensuite été évaluées grâce à l'ensemble AROME. On a pu montrer que les critères d'optimalité pour le filtrage homogène des variances donnaient de très bons résultats, en particulier le critère prenant en compte la non-gaussianité de l'ensemble. La transposition de ces critères à un filtrage hétérogène a permis une légère amélioration des performances, à un coût de calcul plus élevé cependant. Une extension de la méthode a ensuite été proposée pour les composantes du tenseur de la hessienne des corrélations locales. Enfin, les fonctions de localisation horizontale et verticale ont pu être diagnostiquées, uniquement à partir de l'ensemble. Elles ont montré des variations cohérentes selon la variable et le niveau concernés, et selon la taille de l'ensemble. Dans une dernière partie, on a évalué l'influence de l'utilisation de variances hétérogènes dans le modèle de covariances d'erreur d'ébauche d'AROME, à la fois sur la structure des covariances modélisées et sur les scores des prévisions. Le manque de réalisme des covariances modélisées et l'absence d'impact positif pour les prévisions soulèvent des questions sur une telle approche. Les méthodes de filtrage développées au cours de cette thèse pourraient toutefois mener à d'autres applications fructueuses au sein d'approches hybrides de type EnVar, qui constituent une voie prometteuse dans un contexte d'augmentation de la puissance de calcul disponible. / Data assimilation aims at providing an initial state as accurate as possible for numerical weather prediction models, using two main sources of information : observations and a recent forecast called the “background”. Both are affected by systematic and random errors. The precise estimation of the distribution of these errors is crucial for the performance of data assimilation. In particular, background error covariances can be estimated by Monte-Carlo methods, which sample from an ensemble of perturbed forecasts. Because of computational costs, the ensemble size is much smaller than the dimension of the error covariances, and statistics estimated in this way are spoiled with sampling noise. Filtering is necessary before any further use. This thesis proposes methods to filter the sampling noise of forecast error covariances. The final goal is to improve the background error covariances of the convective scale model AROME of Météo-France. The first goal is to document the structure of background error covariances for AROME. A large ensemble data assimilation is set up for this purpose. It allows to finely characterize the highly heterogeneous and anisotropic nature of covariances. These covariances are strongly influenced by the topography, by the density of assimilated observations, by the influence of the coupling model, and also by the atmospheric dynamics. The comparison of the covariances estimated from two independent ensembles of very different sizes gives a description and quantification of the sampling noise. To damp this sampling noise, two methods have been historically developed in the community : spatial filtering of variances and localization of covariances. We show in this thesis that these methods can be understood as two direct applications of the theory of linear filtering of covariances. The existence of specific optimality criteria for the linear filtering of covariances is demonstrated in the second part of this work. These criteria have the advantage of involving quantities that can be robustly estimated from the ensemble only. They are fully general and the ergodicity assumption that is necessary to their estimation is required in the last step only. They allow the variance filtering and the covariance localization to be objectively determined. These new methods are first illustrated in an idealized framework. They are then evaluated with various metrics, thanks to the large ensemble of AROME forecasts. It is shown that optimality criteria for the homogeneous filtering of variances yields very good results, particularly with the criterion taking the non-gaussianity of the ensemble into account. The transposition of these criteria to a heterogeneous filtering slightly improves performances, yet at a higher computational cost. An extension of the method is proposed for the components of the local correlation hessian tensor. Finally, horizontal and vertical localization functions are diagnosed from the ensemble itself. They show consistent variations depending on the considered variable and level, and on the ensemble size. Lastly, the influence of using heterogeneous variances into the background error covariances model of AROME is evaluated. We focus first on the description of the modelled covariances using these variances and then on forecast scores. The lack of realism of the modelled covariances and the negative impact on scores raise questions about such an approach. However, the filtering methods developed in this thesis are general. They are likely to lead to other prolific applications within the framework of hybrid approaches, which are a promising way in a context of growing computational resources.

Title: Ensemble Kalman filter on high and infinite dimensional spaces Author: Mgr. Ivan Kasanický Department: Department of Probability and Mathematical Statistics Supervisor: doc. RNDr. Daniel Hlubinka, Ph.D., Department of Probability and Mathematical Statistics Consultant: prof. RNDr. Jan Mandel, CSc., Department of Mathematical and Statistical Sciences, University of Colorado Denver Abstract: The ensemble Kalman filter (EnKF) is a recursive filter, which is used in a data assimilation to produce sequential estimates of states of a hidden dynamical system. The evolution of the system is usually governed by a set of di↵erential equations, so one concrete state of the system is, in fact, an element of an infinite dimensional space. In the presented thesis we show that the EnKF is well defined on a infinite dimensional separable Hilbert space if a data noise is a weak random variable with a covariance bounded from below. We also show that this condition is su cient for the 3DVAR and the Bayesian filtering to be well posed. Additionally, we extend the already known fact that the EnKF converges to the Kalman filter in a finite dimension, and prove that a similar statement holds even in a infinite dimension. The EnKF su↵ers from a low rank approximation of a state covariance, so a covariance localization is required in...