Application of Artificial Neural Networks to Downhole Fluid Analysis

Abstract

Abstract Reservoir characterization and asset management require comprehensive information about formation fluids. Obtaining this information at all stages of the exploration and development cycle is essential for field planning and operation. Traditionally, fluid information has been obtained by capturing samples and then measuring the pressure/volume/temperature (PVT) properties in a laboratory. More recently, downhole fluid analysis (DFA) during formation testing has provided real-time fluid information. However, the extreme conditions of the downhole environment limit the DFA tools to measuring just a small subset of the fluid properties provided by a laboratory. Nevertheless, these tools are valuable in predicting other PVT properties from the measured data. These predictions can be used in real time to optimize the sampling program, help evaluate completion decisions, and understand flow assurance issues. The petroleum industry has devoted much effort to developing computational methods to model phase behavior. Two approaches are prevalent—simple correlations and equation-of-state (EOS) models. However, in recent years, artificial neural network (ANN) technology has been successfully applied to many petroleum engineering problems, including the prediction of PVT behavior. ANN technology can recognize patterns in data, adjust dynamically to changes, infer general rules from specific cases, and accept a large number of input variables. An ANN architecture can allow for continuous improvement by expanding the training database with new data. In this paper, we present the application of ANN technology to DFA. We demonstrate this with an ANN model that uses the DFA tool measurements of fluid composition as input and produces predictions of gas/oil ratio (GOR), a key PVT property used in real time to monitor a formation tester sampling job. The ANN also provides an uncertainty estimation of its outputs as a quality assurance indicator. We compare ANN results with those from the algorithms used by DFA tools. Introduction Reservoir fluid properties, such as hydrocarbon composition, GOR, density, viscosity, CO2 content, pH, and PVT behavior, are key factors for surface facility design and optimization of production strategies. In most hydrocarbon reservoirs, fluid composition varies vertically and laterally in the formation. Fluids may exhibit gradual changes in composition caused by gravity or biodegradation, or they may exhibit more abrupt changes due to structural or stratigraphic compartmentalization. DFA. DFA techniques, including contamination monitoring, composition measurement, and single-phase assurance, can provide real-time fluid property information during formation testing.1-10 DFA helps ensure that representative samples are obtained and allows an unlimited number of zones to be evaluated in a "fluid scanning" mode. The sampling program can be optimized during the job, and the operator can decide when, where, and how many samples to collect. The ability of focused-sampling cleanup to supply virtually uncontaminated fluids—with faster cleanup time—further ensures optimal DFA results.11,12 Current DFA techniques use the absorption spectroscopy of reservoir fluids in the visible to near-infrared (NIR) region. On the basis of their molecular structure, different types of hydrocarbons have vibration absorptions at different wavelengths, and a simplified hydrocarbon composition can be quantitatively determined from the NIR spectrum. With the latest DFA tool,13 the hydrocarbon composition comprises five groups: methane (C1), ethane (C2), propane to pentane (C3-C5), hexane and heavier hydrocarbons (C6+), and carbon dioxide (CO2). For single-phase assurance it is possible to detect gas liberation (bubblepoint) or liquid dropout (dewpoint) while pumping reservoir fluid to the wellbore, before filling a sample bottle. Despite these advances, DFA measurements are limited by the extreme high-temperature, high-pressure conditions of the downhole environment. Nevertheless, these measurements are valuable in predicting other, nonmeasured PVT properties. They are also valuable in providing uncertainty estimates for the predicted PVT properties. It is desirable to have a means to assess the quality and consistency of the DFA measurements; for example, to determine in real time if a sensor is malfunctioning.