Assessment of different model-management techniques in history matching problems for reservoir modelling

Abstract

History matching is a computationally expensive inverse problem. The computation costs are dominantly associated with the optimisation step. Fitness approximation (proxy-modelling) approaches are common methods for reducing computational costs where the time-consuming original fitness function is substituted with an undemanding function known as approximation function (proxy). Almost all of the applied fitness approximation methods in history-matching problems use a similar approach called uncontrolled fitness approximation. It has been corroborated that the uncontrolled fitness approximation approach may mislead the optimisation direction to a wrong optimum point. To prevent this error, it is endorsed that the original function should be utilised along with the approximation function during the optimisation process. To make use of the original function efficiently, a modelmanagement (evolution-control) technique should be applied. There are three different techniques: individual-based, population- based, and adaptive. By using each of these techniques, a controlled fitness approximation approach is assembled, which benefits from online learning. In the first two techniques, the number of original function evaluations in each evolutioncontrol cycle is fixed and predefined, which may result in an inefficient model management. In the adaptive technique, the number is altered based on the fidelity of the approximation function. In this study, a specific adaptive technique is designed, based on heuristic fuzzy roles; then, for the first time, the applications of all the three techniques are investigated in history matching. To deliver an assessment between the four approaches (the uncontrolled approach and three controlled approaches), a framework is developed in which ECLIPSE-E100 is coupled with MATLAB; and an artificial neural network, a genetic algorithm— with a customised crossover—and a Latin hypercube sampling strategy are used as the proxy model, optimiser, and experimental design method, respectively. History matching is carried out using each of the four approaches for the PUNQS3 reservoir model, while the same amount of computation time was allowed for each of the approaches. The outcomes demonstrate that the uncontrolled approach cannot deliver reliable results in comparison with the controlled approaches, and among the controlled approaches, the developed adaptive technique is more efficient.