AI-Based Decline-Curve Analysis

Manages Reservoir Performance

D ecline-curve analysis (DCA) is one

of the more widely used forms
of data analysis that evaluates well
Step 1—Data Loading and Data-
Quality Checking. Data is loaded into
the platform in a predefined format re-
were to be a sharp change in the produc-
tion behavior.
Step 4—Cross Validation With
behavior and forecasts future well and quired for advanced analytics, machine Real-Life Events. Because the changes
field production and reserves. In the learning, and optimization purposes. The in production are correlated highly with
complete paper, the authors develop input data (including settings and param- real-time events such as perforation
and deploy technologies that apply eters) for technology involve production changes, workovers, compression, op-
DCA methods to wells in an unbiased, data and physical and operational infor- erating hours, and artificial lift, a cross
systematic, intelligent, and automated mation of all wells. validation with the operating condi-
fashion. This method contrasts with Step 2—Machine Learning and tions (if available) is performed to en-
manual DCA, the common practice Smart Clustering Analysis. Cluster- sure robustness and accuracy.
of the industry. ing analysis is performed using machine- Step 5—Curve Fitting. The engine
learning and pattern-recognition tech- computes eventually a decline curve for
Introduction niques to partition the data set into all events using advanced optimization
DCA is used commonly to estimate res- internally homogeneous and externally (minimization, in this case) algorithms
ervoir and well productivity and ulti- distinct groups. These clusters are then for nonlinear least-square problems and
mate recovery and evaluate reserves. used to generate type curves for each selects the best fit for the data while con-
Such analyses are usually performed group (cluster) of wells to apply to the sidering the physical changes in the field.
manually through a curve-fitting pro- wells for which production histories are This not only leads to more reliable pre-
cess by reservoir and production en- not lengthy. dictions but also provides insight to the
gineers using their best judgement Step 3—Event Detection. The next differential estimated ultimate recover-
and experience. Subjectivity is often a step is to identify the last event from ies (EURs) after the corresponding reser-
large component of such estimations, which to apply the decline curve. The voir or operating condition changes.
as are the experience and objectives of event-detection algorithm automati- Step 6—Type-Curve Generation.
the evaluator. cally (without human interference) de- Once all wells are analyzed and their de-
The authors provide an alternative by tects any major changes in production- clines are calculated, the wells in each
developing and leveraging an augmented decline behaviors and flags them as cluster are aggregated to generate type
artificial-intelligence (AI), data-driven a production event. The detection of curves for their respective cluster. These
approach. Geological and engineering such events is crucial in DCA because cluster-based type curves are then ap-
features are considered as well as op- the decline curves fitted before such a plied to the wells in each respective clus-
erational conditions to bring domain break point will not be applicable to the ter with short production histories in
knowledge into the analysis, resulting in post-break-point period. The idea be- order to conduct meaningful DCA.
more-accurate and -reliable predictions. hind the event-detection methodology
In addition, the method is extended to is that, if two linear trends—one with- Probabilistic DCA. Though automated
conduct DCA probabilistically using in a certain interval before the break event detections—augmented with
quantile regression techniques. point (event) and another within a cer- real-life events, robust curve-fitting op-
tain interval after the break point—are timization algorithms, and clustering/
Methods fit, then the difference in trend estimate pattern recognition techniques—can
Deterministic DCA. The work flow is on the break point from the two lin- be used to perform DCA systematically
built upon six steps. ear trends would be significant if there and automatically, the unbiased as-
pect of the process is not fully assured
because the decline is heavily depen-
This article, written by JPT Technology Editor Chris Carpenter, contains highlights of
dent on the choice of the curve-fitting
paper SPE 197142, “Artificial-Intelligence-Based, Automated Decline-Curve Analysis objective function. Depending on the
for Reservoir Performance Management: A Giant Sandstone Reservoir Case Study,” importance of a subset of data points,
by Amir Kianinejad, Rami Kansao, and Agustin Maqui, Quantum Reservoir an infinite number of decline curves can
Impact, et al., prepared for the 2019 Abu Dhabi International Petroleum Exhibition be obtained, with some being pessimis-
and Conference, Abu Dhabi, 11–14 November. The paper has not been peer reviewed. tic, some close to the median, and some

For a limited time, the complete paper is free to SPE members at www.spe.org/jpt.

58 JPT • SEPTEMBER 2020

 EUR

EUR (MSTB)
EUR (MSTB)

1956 1971 1984 1999 2012

1950 1960 1970 1980 1990 2000 2010
Active Shut-in

Fig. 1—EUR estimation of the wells per year, log scale (left) and oil EUR vs. time (right).

optimistic. This issue can be alleviated removed any possible engineer biases. In A key feature of the authors’ DCA tech-
by choosing to minimize the residual a matter of minutes, the tool integrated nology is its machine-learning solution
sum of squares. However, one could all production data and real-life informa- that allows users to partition a data set
argue that some data points are more tion throughout the entire history and into groups that are homogeneous inter-
informative than others. provided a comprehensive analysis at the nally and distinct externally. The clas-
Therefore, the method is extended to a well and field levels. Overall, for this case sification is carried out on the basis of
probabilistic approach using quantile re- study, the spread between the low and a measure of similarity or dissimilarity
gression. This solution allows obtaining the high decline rate has varied from ap- within and between the groups. Different
different confidence intervals for decline proximately 1 to 10%, depending on the cluster features may be selected or can be
curves, thus providing a confidence band amount of historical production. identified automatically. On the basis of
on the declines for individual wells and Fig. 1 shows two metrics derived from those features, the DCA engine clusters
for the entire field. the automated DCA engine used by the the wells and uses them for enhanced
Quantile regression is a method that authors, which helps to understand the type curving.
estimates the conditional quantile or per- reservoir situation better. For instance, The results from the DCA allowed
centile of the response variable on the the left plot shows the average estimat- quantification of the remaining reserves
basis of data and can help obtain a more- ed ultimate recovery per well by the by reservoir with much higher accuracy.
comprehensive understanding of the re- year the well commenced production. They also highlighted multiple potential
lationship between variables of inter- The figure shows that the average EUR areas for improvement to increase the
est. The authors use quantile regression per well has been decreasing since the ultimate recovery of the reservoir.
to simulate the effect of uncertainty in early stages of the field. The well-by-well
a production forecast. The authors’ ap- EUR per spud year (right plot) indicates Conclusions
proach with regard to both well-level and the same trend of decreasing EUR over The main features of the technology
field-level DCA is discussed in detail in time. These metrics highlight some defi- include the following:
the complete paper. ciencies in pressure depletion over time. ◗ Fully automated DCA solution
Another significant benefit of access that allows focus on result
Case Study: to well-by-well decline rates for such a interpretation and analysis
A Giant Sandstone Reservoir large field is that different groupings ◗ Advanced event-detection
The complete paper presents a case study can be analyzed, providing insight into algorithm with a cross-validating
in which the authors’ approach was ap- the performance of different wells of in- framework to compare
plied to a giant mature sandstone res- terest. For example, groupings may be automatically detected events
ervoir in the Middle East. The field has created in terms of well type (vertical, against real-life events
nearly 70 years of production and more deviated, or horizontal), artificial lift, ◗ Flexible machine-learning-based
than 2,000 wells. geographical area, or pads. For this res- solution for precise type curving
Traditional DCA of this field can take ervoir, the different well types exhibit in wells with limited production
several weeks at least if it is performed very similar decline rates (within 1%), history
on a well-by-well basis. The work can yet, not surprisingly, horizontals have ◗ Uncertainty quantification
be sped up if it is split among multiple the lowest declines. Deviated wells have framework based on quantile
engineers, but then different opinions the highest decline rates, possibly the regression providing a confidence
will exist regarding what events are con- result of sidetracks arising to counter interval range for the decline
sidered and where to start the decline. drilling problems. Most of the wells in estimation
The described method dramatically the field are vertical, thus representing ◗ Scalable algorithm to generate
decreased the amount of time required the highest cumulative oil production field, group, and cluster baseline
to analyze such a large data set and among the groups. declines JPT

JPT • SEPTEMBER 2020 59