30 April 2020
14:00 Doctoral defense Fully distance
Theme
Learning-from-Data Approaches to the History-Matching Problem
Student
Cristina Célia Barros Cavalcante
Advisor / Teacher
Anderson de Rezende Rocha (supervisor), Denis José Schiozer (co-supervisor) and Célio Maschio (co-supervisor)
Brief summary
History adjustment is an important reservoir engineering process in which uncertain values of a reservoir model are changed in order to find models that honor historical production data. It is a typical reverse and poorly posed problem, which admits multiple solutions and plays a fundamental role in reservoir management tasks: reservoir models support strategic decisions in the development of the field, and the better they are calibrated, the greater the confidence in their predictions about the real performance of the reservoir. Although the literature on historical adjustment has made remarkable progress in the past two decades, given the uniqueness of each problem, no strategy is proven to be effective in all cases, and research efforts to find alternative methodologies are always invaluable. This thesis aims to investigate the adequacy of learning approaches based on data for the problem of historical adjustment, verifying whether the use of data from available solutions can guide, in an efficient and effective way, the adjustment process towards solutions of better quality and with reliable forecasting capacity. Four data-oriented methods have been proposed and implemented that, using machine learning and / or optimization techniques, continuously discover, among the available solutions, the patterns of input attributes that lead to good output responses, and that can be used in generation of new solutions. The design of these methods involved the proposal of a software architecture based on reusable components to deal with needs typically present in historical adjustment processes such as consultation and analysis of available solutions, partitioning the grid of the reservoir model in regions of interest and manipulation of uncertain petrophysical or global attributes. In all methods, information learned from the dynamic assessment of available solutions supports strategic decisions about what needs to be changed, and where and how changes should occur, in order to generate new (and hopefully better) solutions. The proposed approaches were validated using the UNISIM-IH benchmark, a challenging synthetic case based on the Namorado field, Campos basin, Brazil. The results indicate the potential for learning from the data to generate multiple solutions that not only honor historical data but, above all, have acceptable performance when predicting reservoir production. Compared to historical adjustment methodologies previously applied to the same benchmark, the proposed approaches have competitive results in quality and / or forecasting capacity, and require substantially fewer simulations.
Examination Board
Headlines:
Anderson de Rezende Rocha | IC / UNICAMP |
João Paulo Papa | FC / UNESP / BAURU |
André Carlos Ponce de Leon Ferreira de Carvalho | ICMC / USP |
André Ricardo Fioravanti | FEM / UNICAMP |
Alessandra Davólio Gomes | CEPETRO / UNICAMP |
Substitutes:
Sandra Eliza Fontes de Avila | IC / UNICAMP |
Hélio Pedrini | IC / UNICAMP |
Jefersson Alex Dos Santos | DCC / UFMG |