Quality Assurance Tool for PVT Simulator Predictions
- N. Varotsis (Schlumberger) | V. Gaganis (Consultant) | J. Nighswander (Schlumberger)
- Document ID
- Society of Petroleum Engineers
- SPE Reservoir Evaluation & Engineering
- Publication Date
- December 2002
- Document Type
- Journal Paper
- 499 - 506
- 2002. Society of Petroleum Engineers
- 5.2.1 Phase Behavior and PVT Measurements, 4.6 Natural Gas, 5.6.4 Drillstem/Well Testing, 5.6.1 Open hole/cased hole log analysis, 4.1.9 Heavy Oil Upgrading, 5.5 Reservoir Simulation, 5.1 Reservoir Characterisation, 1.14 Casing and Cementing, 4.1.5 Processing Equipment, 4.1.9 Tanks and storage systems, 4.1.2 Separation and Treating, 7.6.6 Artificial Intelligence, 2 Well Completion, 6.1.5 Human Resources, Competence and Training, 5.2 Reservoir Fluid Dynamics, 5.2.2 Fluid Modeling, Equations of State
- 1 in the last 30 days
- 569 since 2007
- Show more detail
- View rights & permissions
|SPE Member Price:||USD 12.00|
|SPE Non-Member Price:||USD 35.00|
The currently available pressure/volume/temperature (PVT) simulators predict the physical properties of reservoir fluids with varying degrees of accuracy depending on the type of model used, the nature of the fluid, and the prevailing conditions. Nevertheless, they all exhibit the significant drawback of lacking the ability to estimate the quality of their answers.
Artificial Neural Networks (ANNs) trained by large PVT databases are increasingly used to provide accurate predictions of physical properties mainly because of their ability to learn from experience. The use of such models offers the unique capability of estimating the quality of their predictions, as the degree of competence can be evaluated for each unknown test case. The accuracy of the ANN-based PVT simulators depends heavily on the density of the database compositional mapping around the coordinates of the unknown reservoir fluid. Unknown test cases found outside the available training space may lead to poor predictions.
In this work, a quality assurance tool is presented that is integrated to the PVT Expert,* which is an ANN-based PVT properties prediction model. In the case of an unknown fluid, this tool qualifies the predictions of the PVT simulator based on the evaluation of the affinity of the test case with the training data sets contained in the database used. Subsequently, the competence with which the ANN model has learned the general trend in the area around any new test case is assessed numerically through the development of one ANN sextuple model per property. Finally, the use of the average predicted value eliminates the risk of considerable deviations.
This innovative approach was tested successfully against a large set of studies unseen by the ANN model. The ability to provide confidence for the accuracy of the PVT predictions and to assess their quality significantly upgrades the applicability of the PVT simulator as a valuable reservoir management tool.
An ANN can be defined as a multidimensional function that includes a large number of parameters relating input and output data.1 Advantages such as its ability to learn the behavior of a database population by self-tuning its parameters, the performance of direct and rapid calculations, and its capability of becoming increasingly expert by retraining render this tool suitable for applications such as the prediction of PVT properties.
The application of ANN techniques is rapidly gaining popularity in petroleum engineering problems such as reservoir characterization, 2,3 well completion and cementing,4 well-test interpretation, 5 well logging,6 reservoir simulation,7 and prediction of equilibrium k-values.8 Recently, ANNs were used to predict the bubblepoint pressure and formation oil volume factor of crude oils given the reservoir temperature, oil API gravity, gas/oil ratio (GOR), and gas specific gravity.9
The PVT prediction model that was presented in our previous work10 is an ANN-based simulator that has been developed with a database containing the PVT data of 650 reservoir fluids from around the world measured in the laboratory. These data cover the complete range of fluid types and compositions as well as the reservoir operating conditions, with approximately 400 of the data sets belonging to reservoir oils and 250 data sets referring to gas condensate fluids.
The type of PVT properties considered and the range covered in this data set are summarized in Table 1. Although each PVT property in this data set varies over a wide range, the data population is not uniformly distributed and, as expected, is biased toward low-volatility oils and gas condensates of low liquid yield. To account for this data bias and to achieve a uniform performance of the ANN models, appropriate nonlinear transformations were selected and used.
Key measurements that could be performed rapidly on the produced fluids, either on site or in the PVT lab, were selected as input data for the developed ANNs. For reservoir oil fluids, the input data set consists of reservoir fluid composition, saturation pressure, reservoir temperature, fluid density at the bubblepoint, viscosity and density of stock-tank oil, and flash gas density. The input for gas condensates consists of reservoir fluid composition, dewpoint, reservoir temperature, z factor at the dewpoint, field GOR, and tank liquid density.
The model provides the full set of properties obtained from a full PVT study, including the constant mass study (CMS); the differential vaporization (DV) or the depletion study (CVD); the separation test (SAST); and the viscosity study (VS). Curves such as the isothermal compressibility, the oil volume factor, the GOR, the viscosity, the oil-phase density, the gas-phase z factor, the cumulative produced fluid, and the retrograde liquid deposit are predicted in the full pressure range.
To allow the ANNs to predict full-pressure-range PVT properties for each fluid with sufficient accuracy, the data curves were split into monophasic and diphasic regions. The remaining monotonic curves were further normalized in the unitary square. This corresponding states-like approach attributes similar shapes to the transformed physical property curves of reservoir fluids exhibiting similar thermodynamic behavior, even if the actual values of the original curves differ. For example, by normalizing the diphasic Bo curve in a unitary square, the entire set of Bo curves is included between the curve shape types of the very low- and very highvolatility oils (Fig. 1). Consequently, the transformed curves obtained are discretized to obtain a set of numbers that can be reproduced by an ANN. The curve endpoints are predicted by singleoutput ANNs. The transformed curves proved to be much easier for the ANNs to learn than the original ones. It is noted that special inversible, nonlinear transformations are required to turn curves such as the liberated gas specific gravity into monotonic ones.
We found that some of the single-valued properties could be associated with a known measurement. For example, the reservoir oil density at atmospheric pressure is related to the measured density at the bubblepoint. Instead of aiming directly at the original property, the difference (or the ratio) of its value with respect to a related property value is predicted. Using this technique, even large deviations of the difference/ratio predictions result in only small errors in the original property.
|File Size||821 KB||Number of Pages||8|