SPE 116162

PVT Data Quality: Round Robin Results A.G. Stephen, SPE, D. Bergman, SPE, and T. Dodd, SPE, BP, and W. Kriel, SGS

Copyright 2008, Society of Petroleum Engineers This paper was prepared for presentation at the 2008 SPE Annual Technical Conference and Exhibition held in Denver, Colorado, USA, 2124 September 2008. This paper was selected for presentation by an SPE program committee following review of information contained in an abstract submitted by the author(s). Contents of the paper have not been reviewed by the Society of Petroleum Engineers and are subject to correction by the author(s). The material does not necessarily reflect any position of the Society of Petroleum Engineers, its officers, or members. Electronic reproduction, distribution, or storage of any part of this paper without the written consent of the Society of Petroleum Engineers is prohibited. Permission to reproduce in print is restricted to an abstract of not more than 300 words; illustrations may not be copied. The abstract must contain conspicuous acknowledgment of SPE copyright.

Abstract

Compositional data is a vital input to many engineering models, including the calculation of the mud free properties of reservoir fluids sampled by repeat formation testers and calculating the value of gas sold on a calorific value basis. For all such applications, the quality of compositional data is of critical importance. The impact can be material in a wide range of issues, including reserves estimation, facilities design, gas hydrates prediction and sales gas valuation. The main aim of this study was to set quantitative criteria for screening laboratories prior to tendering for a PVT contract.

Reliance on contractor quality assurance procedures is not sufficient. Routine internal consistency checks may not

identify some of the errors. Evaluation based on analyses of identical field samples relies on one data set being of high quality, an assumption that may not be correct. There is a need for an independent supplier of high quality gas, liquid and live fluid samples of accurately known compositions that can be used by laboratories to demonstrate the quality they can achieve to the oil companies before unique and valuable samples are sent to them for analysis.

The results of a Round Robin evaluation of PVT laboratories around the world are presented. A set of identical

samples was sent to each laboratory 3 dry gases, 3 stock tank liquids and one live sample. There was considerable variation in the quality of compositional analyses reported. A small minority of the laboratories tested generated compositional data that fully met the highest quality measures for both gas and liquid analysis. A similar number were close to meeting these standards, but the majority generated data that was deeply flawed in some respects. In one case, interpretation errors resulted in 24% reduction in C4+ in all the gases. The resulting calculated calorific values were as much as 5% low a large potential loss of value if used to calculate the value of sales gas. All laboratories were given feedback, and some are actively engaged in resolving the problems identified.

Introduction

Laboratory measurements play a critical role in many industries, and where the information provided has significant financial or safety implications many of these industries have implemented clear procedures not only for measurements but for quality control and assurance. Examples include Gold assay, where the purity of the gold established in the laboratory has a direct relationship to its price, medical laboratory analyses, where the data quality is literally a life and death issue and refrigerant viscosity (sale et al, 2000) where the device performance is crucially dependant on the fluid properties. There is increasing application of international standards for laboratory quality, particularly where the quality of the data reported is audited by an external, independent body for these businesses. In the oil industry, PVT laboratories provide data vital data and are often accreditation to international standards that are subjected to external audit, but the protocols relate to management systems and do not provide inputs or insights into the quality of data being generated merely that procedures are written down and followed. Each laboratory develops its own experimental procedures, quality control procedures and establishes its own target quality limits. Whilst the individual laboratory may meet its own specifications routinely, in cases where the criteria are inadequate or insufficiently challenging, the result can be data that is sufficiently inaccurate as to have a material impact on the value of a project.

Value of PVT data PVT laboratories generate data that is used in many applications throughout field life, including reservoir

simulation, reserves submissions, facilities design, well bore and flow line dynamics, surveillance and fiscal allocation. The value of such data is best understood in relation to specific examples where the data did not meet minimum quality

2 SPE 116162

expectations. In one example, the price of sales gas was set by the calorific value of the gas composition as determined by gas chromatography. The methane content of the gas was 93mol%, but there was a 3mol% variation in the analysis. Had this error not been identified, the impact over field life would have been of order $50million. In a different example, a lean gas composition was being used to test reservoir connectivity. CO2 analyses varied between 3 and 4 mol% - which was found to be due to poor quality control procedures. Undetected, this would have impacted well density and the associated gas line infrastructure. In another example, oil samples collected using MDT samplers were flashed and analysed to be used in an equation of state based phase behaviour model to calculate the mud free reservoir fluid properties. The laboratory was using a technique in which flashed gas was recycled into the flashed oil to ensure that the gas and oil were at equilibrium and homogenised. The temperature of the system was not constant throughout, resulting in a condensate drop out at some point in the gas line that was not detected. This led to the loss of samples that cost around $1mm to capture in rig time alone, but had this been a smaller field with fewer opportunities to appraise it, there would have been a direct impact on reserves, predicted well rates and facilities design.

PVT contract tenders

The greatest opportunity to minimise the risk of loss of value due to defective data is in the early planning stages, by ensuring that the clearly defined data quality measures can be achieved before a laboratory is selected. Often, this is done on a study by study basis, where the value of the contract is relatively small. However, there are instances where a large planned appraisal programme is rolled up into a single contract. Under these circumstances, formalised contracting procedures are called for, in which any screening criteria, whether technical, commercial or safety, are fully specified before the tendering process begins. A method to provide an objective test of laboratory data quality was developed and is reported here, although the method was used with more laboratories than were included in the original scope of work.

Options for screening laboratories for PVT contracts Two main options were available either to rely on the laboratories own quality control measures, or to provide samples for analysis that would allow an objective review of the accuracy of measurements being offered.

Rely on the laboratories quality systems The extreme examples of defective data were chosen to illustrate the need for care in all stages of data acquisition

and interpretation. They also illustrate the risks inherent in relying on the simplest method for choosing a laboratory simply relying on self assessment. There are few universal standards that define quality in PVT and associated measurements though there are some well documented and robust procedures for gas analysis (Beaty et al, 2005) as well as a commercially available liquid standard (Phillips Gas Oil standard) that can be used. The result is that laboratories adopt their own procedures, quality assurance and quality control measures but many are not extensive enough, allowing the potential for significant errors to pass unnoticed.

Provide reference samples for an objective assessment of data quality offered

There are many cases where there is good reason to need measurements from an unfamiliar laboratory. This can be a new facility, or more commonly, an existing laboratory in a new area of exploration. An approach that has been used many times to explore the capability of such unfamiliar laboratories is to send duplicate sets of samples to the local laboratory and to one that has a long track record of high quality (Bergman 2001, Stephen, 2007). This can be a very powerful technique, as it includes routine quality control checks for internal consistency in the data provided by each laboratory, but also allows direct comparison between the reported data. This has been used to provide clear feedback to laboratories, but none of this has been published. The approach has some limitations, including the concern of the unknown laboratory about the validity of the assertion that the reference laboratory is indeed producing high quality data, as well as only providing an evaluation of the laboratory capability after the first sample has become available, rather than before it.

In 1990, round robins on black oil and gas condensate measurements were organised (Merrill, 1990, Sawdon,

1989), in which a set of identical samples were sent to a number of laboratories for analysis. These were recombined to generate reservoir fluids that were then used for basic PVT studies. This approach is relatively simple to carry out, and provides useful data on inter-laboratory variations in measurement that has provided an insight into measurement accuracy. This method leaves uncertainty about various issues, including whether poor GOR measurements are due to incorrect recombination ratios, leaking shipping vessels or mis-measurement in the laboratory. More fundamentally, the method uses the averaged result from all laboratories to define the average value for comparison. This can yield misleading conclusions where some laboratories have similar deficiencies in measurements. An example of this would be poor handling of flashed gas and liquids, leading to a bias towards low reported concentrations in the ethane to nonane region. Laboratories that overcome this problem can be penalised by apparently reporting excess material relative to the average. Alternatively, laboratories that have a consistent track record can be used to provide a reference analysis, but this is a subjective approach and is open to challenge.

SPE 116162 3

A better approach is exemplified by the Gas Processing Association (GPA) that ran a Gas analysis round robin, (sale, 2000), in which gas mixtures were manufactured by adding individual components into a mixture and weighing the container before and after each addition. This gave certified compositions that were independent of the analysis method used by the laboratories being evaluated, to an accuracy at least as good as that of the analyses methods being evaluated. Adopting this approach for the PVT round robin samples would eliminate all the key concerns over the comparative approach used previously. The priority for this work was to provide a clear measurement of the quality of data received from laboratories, with the main emphasis on compositional data. It was therefore decided to develop an approach to rigorously test the reservoir fluid composition. The fluids were designed to test the accuracy of gas analysis using dry gases to ensure that any errors in the gas analysis were due solely to chromatography, with no contribution from sample handling. A range of gas compositions were needed, both to provide verification over a wide range of CO2, ethane and methane contents, but also to identify any systematic biases. Liquid samples that had been heated and filtered ensured a similar test of chromatography for liquids, by ensuring that no further weathering would occur Preparation of samples

Gas mixture preparation

Three gas samples were by a commercial speciality gas manufacturer by weighing a pressure vessel before and after each pure component was added, and verifying the certified composition by GC analysis. Three gas mixtures were needed to provide a range of concentrations of methane, CO2 and ethane, to test the linearity of calibration over a wide range of compositions. All of the gas mixtures were designed to be dry gases so that any defects could only be due to calibration or operation of the GC instrument. Quality control of the samples provided included analysis by gas chromatography by the coordinating laboratory. This provided a cross check on the certified composition, to guard against gross errors, and to verify that the sub samples were identical. These analyses were not used at all in interpreting the quality of data received. Gas A had the highest proportion of heavier components, but it was predicted to be a dry gas. The routine QC analysis of the gas A sub samples showed that the gas had taken the gas below its dew point. Investigations showed that the pressure of the bulk sample of Gas A indicated a slight leak during transit. There was insufficient time to get another gas made up, so the options were to reject gas A or to send it out as received. As all sub samples were reported as being identical within the precision of the QC analysis checks used the decision was taken to ship the samples to the laboratories, and recognise that the composition of this gas was not that certified by the manufacturer. Instead, its composition would be determined by averaging the analyses of gas A from those laboratories with the highest quality data of gas mixtures B and C.

Liquid mixture preparation

Preparation of a liquid by weighing in pure components would have provided a definitive composition that would have tested only the calibration of the response factors for the different components. This would have left untested the more challenging issues of integration of the unresolved isomers that dominate the liquid analysis beyond about C8, and would also have left untested the accuracy of determination of the proportion of the liquid that has too low a volatility to reach the detector the plus fraction. These factors require a real oil sample to be used, but this makes it very much more difficult to define the reference composition robustly.

A stock tank oil sample was heated to 30C to weather it to ensure that there was no possibility of further loss of light

components during shipping, sub sampling or analysis, thereby retaining the integrity of the test. It was also filtered to remove any solids to avoid any concerns about segregation during storage and shipping. A sub sample was distilled, and the distillate cuts analysed by gas chromatography. The plus fraction that is not detected by GC methods was concentrated in the higher distillate cuts, reducing the uncertainty in the C36+ compared with GC analysis of the whole sample. GC analysis of distillate cuts can lead to large errors at each end of the cut, due to small errors in the definition of the end point of the carbon number pseudo components, an effect that is illustrated by the mud base oil mixture in this study. However, the overall weight distribution is fixed by the weight fraction of each of the cuts, and these GC errors lead only to fluctuations within the cut ranges. For the purpose of this work, then, the C35- distribution was determined by averaging the results of the best laboratories that showed no overall trends with carbon number, as verified by the distillation.

Being unable to prepare the liquid sample gravimetrically represents a potential weakness in the evaluation, despite

the supporting evidence from distillation. It was decided to extend the test by weighing in a proportion of a mineral oil as used in oil based drilling mud. Two such mixtures were made, with 5.5 wt% and 13.8 wt% base oil in the mixtures. These were chosen as being representative of many samples received. A sample of the base oil was also shipped as a liquid sample, to mimic procedures used for PVT studies on contaminated samples taken by formation testers. This combination of samples provided a direct check of the accuracy of quantification of the mud contamination in bottom hole samples, a parameter that is vital to the providing information about the reservoir fluids, and as such was a highly valuable measure of laboratory

4 SPE 116162

quality by itself. However, the addition of this as an internal standard into the samples gave a direct quality measure to check the mixtures provided.

As with many base oils, the component range was sufficiently limited that distillation would have provided no useful data. The only way to establish a reference composition for this base oil sample was to use results from the participating laboratories.

Results

Interpretation method

The data received from participating laboratories are shown as percentage deviations from the reference composition. For gas analyses, molar (mol %) compositions are used, and for liquid and live fluids, weight (wt %) compositions are used.

Relative Deviation % = (reported value reference value)/reference value * 100 Based on an evaluation of measurement uncertainties in analyses (Doghmi, 2007), it was asserted that the

uncertainty in each component is the sum of a part that is proportional to the amount of the component present in the reference mixture and an absolute value, to reflect the accuracy of measurement and calibration, but also to include effects such as rounding error and baseline uncertainties. For a component (i) with a reference composition of yi, the allowable error limit of the highest quality band is defined as Ai * yi + Bi. If the actual error was less than the allowable error then a score of 1 was assigned for that component. A score of 2 is given if the actual error is between 1 and 3 times the allowable error. If the actual error was more than 3 times the allowable error a score of 3 was given. The weighted average of these scores was calculated for each fluid, and averaged for the gas and liquids separately. The average of the gas and liquid score gave the overall rating. The values of Ai & Bi used were based on experience of the individual component mass balance errors in experiments such as differential liberation, and on the rare occasions when high quality laboratories had reason to analyse the same samples. The values used for this work are summarised in table 1. For most components, this approach could be used for any sample. However, methane comprises such a large proportion of these gas samples as to present a special case. If methane at 90 mol% recorded a 3 mol% high response, the result would be a composition that summed to 103%. Renormalizing this to 100% would reduce the methane concentration to 93*100/103, i.e. 90.3mol%. Renormalisation has reduced the error in methane from 3 % to 0.3 mol% in this case. For this work, a constant but reduced value was used for methane. The plus fraction, that is, the material insufficiently volatile as to reach the chromatography detector, has different factors contributing to its error, and for this work a single value has been assumed. A different fluid with different C36+ would have very different errors, and care should be taken in selecting an appropriate value for, say, a heavy oil, or a gas condensate.

Focusing on the relative error rather than the absolute errors makes it easier to identify trends in data, but there is a

risk of putting excessive emphasis on large relative errors in small quantities. The same absolute error of 0.09 % for a component at a concentration of 90% represents a relative error of 0.1%, but the same absolute error for a component present at a concentration of 0.09% represents a 100% relative error. Plotting the boundaries of the defined quality bands allows this effect to be seen, and helps prevent over emphasis on small errors in trace components.

Gas mixtures Most of the laboratories tested provided good data for all the gas analyses, but there were exceptions. . The most

clearly identifiable error was that of the gas analyses from Lab E, which all show an abrupt change of 24% loss from iC4 onwards (Figure 1). The laboratory had made an error in tying the data from the two detectors in the gas analysis, and removing this mistake led to high quality gas analyses.

Liquids 1 & 2 Fewer laboratories reported high quality data for these liquid analyses. Several laboratories show a roughly constant

relative deviation in the C5 to C35, with a compensating error in the C36+. This pattern frequently results from problems with the internal standard method for quantifying the C36+ fraction. An example of a laboratory that appears to have used an inappropriate integrations technique by defining the baseline by the valleys in the signal is illustrated in Figure 2. The impact of such misinterpretation is illustrated along with the results from the best liquid analyses included for comparison in Figure 3. Plotting the quality metrics for gas vs. liquid analyses (Figure 4) shows no correlation between the quality of the liquid analyses and the quality of gas analysis.

SPE 116162 5

Mud contamination estimates Liquids 1 and 2 were mixtures of base oil and a stock tank oil, allowing the impact of liquid analysis quality on mud

contamination estimates to be checked. The results are shown in Figure 5 as absolute deviations from the gravimetrically determined base oil content of each of the two liquid samples as a function of the quality metric of the liquid analyses. The most extreme errors (+11wt%,-20wt %) are off scale on this chart. The best results for mud contamination were obtained from those laboratories that produced the highest quality liquid analyses. In one case, the liquid errors were very high outside of the mud component range. By coincidence, these errors cancelled in the mud component range, leading to a surprisingly accurate mud contamination level for both liquid. In all other cases, there was a good correlation between liquid analysis quality and the accuracy of mud contamination.

Discussion

Some laboratories fully met the accuracy expectations implied by this work, yet not all did so. It may be that one reason that this situation has developed is a poor understanding of just how wide the range of quality on offer is, as well as a poorly understood relationship between data quality and business performance. The criteria used in this work were based on a view of what could be achieved by a good commercial laboratory, but not on any direct evaluation of the consequences for decision making. One of the business issues that motivated this work was the poor analysis of gas samples intended to calculate the price of sales gas based on calorific value. For the gas analysis data grouped in the highest quality band, the average error on CV was -0.01% with a standard deviation of 0.6%. The lowest quality data group had a maximum error of -5%, an average error of -1% and a standard deviation of 2%. If such errors from the worst data set remain undetected, the revenue stream from gas production would be reduced by the bias evident in the data.

Each laboratory was given feedback on the quality of their data including, where appropriate for those laboratories

that did not provide data in the highest quality band, an evaluation of the most likely cause of errors. The nature of the errors was varied, but there is one common contributory factor. If the laboratories method development and ongoing quality assurance/quality control methods had included stringent tests of capability using realistic and well characterised fluids, the flaws identified by this work would have been apparent to the laboratories. This was not the case, and to their credit, several of the laboratories concerned are working cooperatively to improve performance.

To understand the impact of the compositional errors on fluids modelling, an approach was used in which an

equation of state model for a gas condensate and a black oil model were used to predict the saturation pressures based on the reported fluid compositions. The equation of state calculated flashed gas and liquid compositions which were used to defining the reference compositions, along with the calculated molar ratio of flashed gas and liquid. For each laboratory, the flashed compositions were scaled according to the results of their analyses of gas B and liquid 1, and these adjusted compositions used to calculate the resulting reservoir fluid composition which was then used with the same equation of state models to predict the saturation pressures (Figure 6). The data in the highest quality band all resulted in very good agreement with the reference case. By contrast, the lower quality data shows considerable variation underestimating the dew point of the gas condensate by over 50 % in some cases. The errors in predicting bubble points were as large as 20%. Although the bubble point predication is less sensitive to compositional errors than gas condensate dew points are, the magnitude of the error in prediction is very large, and would lead to significant errors in depletion planning and in facilities design.

During the course of this work some procedural issues were overcome that a successful and reliable shipment of

gravimetrically prepared live oil samples for a second series of round robin evaluations. This allows this evaluation work to be extended to include sample handling quality. Evaluation of physical property measurements, however, can only be done on the basis of comparisons between laboratories.

For this work, the uncertainty, i, was characterised by Ai * yi + Bi, where Ai and Bi are constants defined for each

component. In the case of a 3% uncertainty in methane concentration in a gas with 90% methane, the reported methane content would be 93% and the composition would sum to 103, Renormalisation to 100% by applying a factor of 0.97 reduces the reported methane content to 90.29% - an error of 0.29%. This renormalisation effect should be included in the interpretation of data in the future, in order to allow the method to be more generally applied, using the same coefficients for almost all components:-

i= (Ai * yi + Bi)*100/ (100+ Ai * yi + Bi)

Conclusions The methods used to prepare, define and check the reference compositions for the dry gas mixtures were effective.

The addition of known quantities of mud base oil to a stabilised stock tank oil that had been analysed by distillation allowed the reference compositions to be defined with confidence.

6 SPE 116162

Whilst some laboratories provided very high quality data (27% of those tested), 54% of laboratories tested provided data that was fell into the lowest quality category, with the remaining 18% in a quality band that would allow use for some limited purposes.

The range of errors observed was large o Up to 3% in the methane content in gas o Up to 17 wt% in mud base oil content of stock tank liquid o C36+ wt% errors varied between +11wt% and -20wt%, for a reference value of 20 wt%

The errors observed were generally consistent for the different samples from the same laboratory, showing not only a bias in the methods used, but also insufficient quality control and quality assurance measures were used. Whatever the specific technical issue that caused the errors in individual laboratories, systematic errors can only have been generated where the evaluation of procedures, quality assurance and quality control did not provide an adequate test of the methods used on the type of samples received from oil and gas fields

The probable cause of errors included:- o Flawed procedures. Two laboratories used the incorrect valley to valley integration method

instead of integrating to the known baseline for liquid analysis. o Inaccurate quantification of the low volatility material that is not detected (C36+) o Failure to follow procedures. Calculation of results in a spreadsheet with an error in the formulae,

when the spreadsheet was not part of the measurement protocol and was not used in routine quality control or calibration measurements.

There were some clear regional trends that may reflect the historical evolution of the PVT markets around the world and the availability of appropriately trained staff, but it would be a serious mistake to pre-judge which regions are most likely to provide the best data.

The methods used in this study are proving very effective in helping some laboratories to recognize and overcome technical challenges, and will lead to greater confidence in using unfamiliar laboratories.

The approach used for evaluating gas and liquid samples has proved to be robust and has helped screen laboratories for data quality, providing them with clear and quantitative targets for improvement and a means to demonstrate future improvements in quality. Lessons learnt from the flaws in an attempted pressurized sample have been applied to a second study, with early indications that this has been successful

A modified approach to assessing error limits has been proposed, that allows for the renormalization effect in components present in high concentrations.

There is an urgent need to have external, independent audit and testing of PVT laboratory procedures and quality control measures. An international accreditation such as that offered by ISO 17025 might meet this need.

SPE 116162 7

Figure 1 Single Laboratory gas analyses - relative errors

8 SPE 116162

Figure 2 Illustration of baseline subtraction

SPE 116162 9

Figure 3 Single laboratory liquid analyses - Percentage errors

10 SPE 116162

Figure 4: correlation of gas and liquid quality

SPE 116162 11

Figure 5 Absolute errors in base oil contamination estimates

12 SPE 116162

Figure 6:- Impact of compositional errors on predicted saturation pressures

SPE 116162 13

Tables Table 1:- Gas and liquid analysis scoring parameters

flashed liquid and live fluid

flashed gas analyses

A B weighting factor

A B weighting factor

% wt% % mol% N2 3 0.01 1 2 0.01 1

CO2 3 0.01 10 2 0.01 10 C1 3 0.01 20 0.5 0.01 20

C2-7 3 0.01 5 5 0.01 5 C8-35 3 0.01 1 5 0.01 1 C36+ 20 0.01 10 5 0.01 1

Table 2:- Summary of scoring system for a single fluid Score for individual component

(score for fluid analysed = weighted average of individual component scores)

Measured error