Vous êtes sur la page 1sur 4

Strategy and Risk Management

TNO | Knowledge for business

Probabilistic history matching

Production forecasting with uncertainty quantication


Current practice in history-matching reservoir simulation models is biased, incomplete and awed when quantifying sensitivities and uncertainties. Using the associated production forecasts for investment decisions represents a signicant risk of sub-optimal eld developments. TNO has developed a new method to remedy this. Its probabilistic framework is generic and can be installed in any oil companys IT infrastructure. Interested parties are requested to contact TNO.
200

Figure 1. Norwegian oil production (mill Sm3 year) for 22 major fields, actual vs. forecast.

150

Original prognosis Actual production Prognosis 1990 Prognosis 1993 Prognosis 1996

not use the available information correctly to calibrate their models. Why the deterministic approach of calibrating reservoir simulation models has become obsolete It is common practice to base a hydrocarbon production forecast on a single deterministic description of a reservoir, one that can reproduce the history production and pressure data to within a certain accuracy. The data, however, are generally consistent with many more models. This renders the adopted reservoir model and associated production forecast quite subjective. Moreover, the uncertainty in any prediction cannot be assessed from just one model. In Figure 2, all cumulative distribution curves of future oil recovery represent fully historymatched data points in a synthetic reservoir simulation study (Ref. [1]). All curves are based on exactly the same model and the same data. What makes them different is how the reservoir engineer (subjectively) chose his workflow. Traditional historymatching has been able to obtain any one point as the most likely or base case. The study demonstrates how limited and biased that practice is. Yet, most long-term forecasts are still based on a single history-matched

100

50

0 1970 1980 1990 2000 2010 2020

Cimulative distribution function

Figure 2. Non-unique history matching solutions and subjective workflows result in meaningless deterministic forecasts (Ref. [1]).

0.8

0.6

TNO-2: (5x6)x3 pars Amoco Isotropic Amoco Anisotropic Elf IFP-STM NCC-Oliver NCC-GA NCC-MCMC truth case

0.4

0.2

0.2

0.4

0.6

0.8

1.2

1.4

1.6

Incremental Oil Production after 16.5 years (in million Sm3)

BenO uGE 009 05-2007

Oil companies are quite poor in production forecasting Although it is their core business, oil companies seem rather poor in production forecasting (see Figure 1). This may have many causes, but a major one is that they do

Strategy and Risk Management

Production forecasting with uncertainty quantification

No simple relationship

Forecasting RMS

3 Forecasting with calibrated models carries inherent uncertainty! This needs to be quantified. Best Match

0.2

0.4

0.6

0.8

1.2

1.4

History Matching RMS

Figure 3. Calibrating reservoir models is a statistical problem. reservoir description. In Figure 3, again from a synthetic case study (Ref. [1]), historymatch quality is plotted against forecast quality. Ironically, the best history match yields the worst production forecast, contrary to common belief! The reason is that the statistical nature of history matching and the inherent bias in the history-matching workflow are ignored. TNOs new method: an automatic, probabilistic history-match procedure An automatic probabilistic history-match and prediction procedure has been implemented by TNO. This procedure automatically generates many realisations of the reservoir reproducing the history data with satisfactory accuracy. Using these realisations, predictions are derived and processed into an expectation curve. Most of the theory behind this methodology has been developed and demonstrated on synthetic cases within the Production forecasting with UNcertainty Quantication project (PUNQ) Ref. [1]. The methodology has been successfully applied to a large, heterogeneous, multi-well gas reservoir, and demonstrates that the PUNQ approach is feasible for a real commercial case under the constraints of moderate computer power and tight time limits Ref. [2].

The process starts with a stochastic realisation of the reservoir porosity and the permeability field agreeing with the well data and the spatial variability Optimisation loop: generation of large number of history matched reservoir models given by a Geological and petrophysical variogram. The so-called object Geostatistical simulation Seed function (a Initial reservoir model measure of the mismatch between observed and Reservoir simulation If object function is simulated data) is sufficiently small Calculation of object function minimised in an If object function is not sufficiently small iterative process Optimization by adjusting the reservoir model. Geostatistical simulation New seed The object Adjusted reservoir model function is recalculated with History matched reservoir model the updated reservoir properties, Running and processing of predictions conditioned to Reservoir simulation: Prediction runs the adjusted porosity and Expectation curves permeability values and well data. The loop is Figure 4. Scheme of the automatic, probabilistic history match procedure.

terminated if the object function value is sufficiently small. This procedure is then repeated with different initial stochastic realisations of the reservoir, which produces chains of realisations. The idea Is that each chain will ultimately lead to a well-matched realisation. The resulting realisations of the porosity and permeability fields will match the petrophysical data at the wells and, unlike the regional multipliers parametrisation of traditional historymatching methods, preserve spatial correlations. Prediction runs are then carried out to assess future production performance using these realisations, with expectation curves being determined for the ultimate recovery, recovery factor, etc. The Maximum Entropy method (Ref. [3]) is used to derive expectation curves that are unbiased, based only on a limited number of reservoir models. The workflow of the history-match procedure is shown in Figure 4.

Worst Forecast

Strategy and Risk Management

Production forecasting with uncertainty quantification

probability of exceedence

Thinking probabilistically: not so evident in spite of perceived necessity Traditional deterministic thinking still prevails. Although most companies agree that uncertainty needs to be quantified, they find it difficult to cross the hurdle to a comprehensive probabilistic implementation of their modelling work. For example, during a well attended workshop held by TNO (EAGE 2001), 80% of the participants replied affirmatively to the following question: When determining the range of future recovery, do you 1. use the most likely static model as input to HM, 2. try to find one single HM-ed model, 3. establish the range of future recovery by permutating some reservoir parameters within acceptable HM-range one by one to establish the impact on future recovery? To the following question 50% of participants replied Yes, and 42% No: Were you aware that techniques exist to construct the fully history-matched expectation curve of future recovery? (i.e. each point on the expectation curve represents a history-matched reservoir simulation model, with the quality of the match having a relation to the likelihood of the model-parameter combination).
Question 19 14% 14%

16 14 12 10 RMS 8 6 4 2 0 1 11 21 31 41 Number of reservoir simulations

RMS (BHP) W-1 RMS (BHP) W-2 RMS (BHP) W-3 RMS (BHP) W-4 RMS (BHP) W-5 RMS (BHP) W-6 RMS (BHP) W-7 RMS (BHP) W-8 RMS (BHP) W-9 RMS (BHP) W-10 RMS (BHP) W-11 RMS (WP) W-3 RMS BHP RMS WP RMS TOTAL

Figure 5. RMS (Root Mean Square) convergence during a single optimisation run. Only well W-3 produces water.
Expectation curve Recovery Factor 1.00 0.80 0.60 0.40 0.20 0.00 0.7 0.75 0.8 0.85 Recovery factor 0.9 0.95 1

Figure 6. Expectation curve of the history-matched recovery factor. Mean = 86%, sigma = 3.3%. From various industry sessions, such as the above workshop, it is obvious that further dissemination of the available practical solutions is necessary, hence this brochure. The main acceptance hurdle seems to be a reluctance to use probabilistic methods all the way, in spite of the evident risk of biased, sub-optimal and incomplete reservoir descriptions. The perceived complexity of probabilistic methods among reservoir modellers, and a lack of management pull, are suggested causes. Gradually this attitude seems to be shifting, however. A commercial case study The history-match procedure has been applied to a heterogeneous, compartmentalised gas reservoir with over 10 dynamic wells and 17 years of production history. Using a commercial reservoir simulator, it was modelled by some 20,000 active grid blocks (12 layers), including some 9,300 non-neighbour connections. The two main issues that needed to be resolved in the reservoir study were the estimation of the gas initially in place (GIIP) and the ultimate recovery (UR). In particular, the interested party wished to obtain the history-matched uncertainty range of these two quantities. The PUNQ approach was used to assess these quantities, taking into account the posthistory matching uncertainty in the reservoir properties. The history match was performed on well bottom hole pressures (BHP) and cumulative water production (WP). Figure 5 presents an example of the RMS (Root Mean Square) convergence during a single optimisation run of the individual wells, the BHP and WP RMS and the total RMS. A single optimisation run was ended after 45 reservoir simulation runs as no significant reduction of the object function occurs by then. Approximately one out of three optimisation runs did not converge, so that nearly 2000 reservoir simulation runs were needed to obtain 25 history matches and as many predictions. The duration of an

A B C D

14%

58%

And to the related question: Re previous question, if YES, why does your company not request such techniques to be applied in-house / further developed?. The answers were: A Technique too immature, waiting for R&D B Technique too difficult or too cumbersome C Too little added value D OTHER, please specify

Strategy and Risk Management

Production forecasting with uncertainty quantification

Geo energy and Geo information

optimisation run is about 5 hours on a 1.8 GHz PC. Following the prediction runs, the next step is to determine the probability density functions (pdf) for the GIIP and the UR. Within the chosen workflow we can use the Maximum Entropy method to derive expectation curves that are unbiased, based only on a limited number of reservoir models. Figure 6 shows the expectation curve of the recovery factor. Your opportunity to assess and improve the validity of your reservoir simulation models The PUNQ methodology has been successfully applied to a real commercial case. A series of accurate history matches was obtained that yielded a series of production forecasts. Using the Maximum Entropy method this series was processed into a fully history-matched expectation curve of future recovery. This reservoir study demonstrates that the PUNQ approach has become a practical proposition for real, fullfield commercial cases. In comparison, then, the traditional approach of matching a

single, most likely realisation is no longer recommended because it is no longer a necessity imposed by limited computer power and/or immature research results. TNOs probabilistic framework is generic and can be installed in any oil companys IT infrastructure. Interested parties are requested to contact TNO. References [1] C.F.M. Bos et al., (1999), Production Forecasting with UNcertainty Quantification: PUNQ-2. TNO Report, NITG 99-255-A. [2] P.J.P. Egberts, G.K. Brouwer and C.F.M. Bos, History matching and forecasting with uncertainty quantification; a real case study, paper F-09 presented during the EAGE 64-th Conference & Exhibition, Florence, Italy, 27-30 May 2002. [3] M. Nepveu, (2000), Hydrocarbon Production Forecasting with Limited Data. In: Proceedings of the 7th European Conference on the Mathematics of Oil Recovery. Baveno, September 2000.

TNO Built Environment and Geosciences Geological Survey of the Netherlands is the central geoscience centre in the Netherlands for information and research to promote the sustainable management and use of the subsurface and its natural resources. TNO Built Environment and Geosciences Geological Survey of the Netherlands Princetonlaan 6 PO Box 80015 3508 TA Utrecht The Netherlands T +31 30 256 46 00 F +31 30 256 46 05 E info-BenO@tno.nl tno.nl

Vous aimerez peut-être aussi