Oil and Gas companies are always concerned with continuously evaluating reserves of their assets. Calculating reserves is not an easy task since it requires full knowledge of many technical and non-technical aspects regarding the reservoir nature, available budget, utilized technology, economical conditions and others. Generally, the most important parameter in calculating the reserve for new fields/reservoirs is the “Recovery Factor”. Therefore, many technical approaches available to estimate hydrocarbon reserves are related to estimation of recovery factor.
Recovery factor considers all the aforementioned technical and non-technical aspects. And because these aspects can be numerous, with some that can be only quantified subjectively, the first step in our approach was to standardize and normalize many of these aspects in order to formulate them into digital values to be used in Neural Network algorithms. We gathered all those aspects and classified them into two main categories: Quantitative parameters and Qualitative parameters. Quantitative parameters are normally available in digits (e.g. permeability, porosity, net-to-gross, reservoir pressure… etc.). Nevertheless, this is not the case in handling the Qualitative parameters since parameters like Technology, Asset Remoteness and reservoir structure complexity are not normally represented in digits.
By using wide range of actual fields' data, we had a very good calibration and control points on the algorithm we are going to use. These data include almost all the governing parameters such as Asset Economics, Technology, Facilities, Start of Production, Number of Wells, Reservoir Architecture, Rock and Fluid Properties, Reservoir Energy and others. Since some of these aspects may be hard to find, we generated two well-trained Artificial Neural Network (ANNs): (1) Simple NN includes a few “easy-to-know” basic data, and (2) Sophisticated NN that includes many advanced data with reasonable accuracy. These NNs were taught with more than 150 lessons at which the inputs (lessons) were the aforementioned “parameters” and the output was the “Recovery Factor”.
Testing both the simple and sophisticated networks showed prediction capabilities of 9.5% and 8.0% of actual recovery factor, respectively.