AI/machine learning

AI-Based Decline-Curve Analysis Manages Reservoir Performance

Decline-curve analysis is one of the more widely used forms of data analysis that evaluates well behavior and forecasts production and reserves. This paper presents technologies that apply DCA methods to wells in an unbiased, systematic, intelligent, and automated fashion.


Decline-curve analysis (DCA) is one of the more widely used forms of data analysis that evaluates well behavior and forecasts future well and field production and reserves. In the complete paper, the authors develop and deploy technologies that apply DCA methods to wells in an unbiased, systematic, intelligent, and automated fashion. This method contrasts with manual DCA, the common practice of the industry.


DCA is used commonly to estimate reservoir and well productivity and ultimate recovery and evaluate reserves. Such analyses are usually performed manually through a curve-fitting process by reservoir and production engineers using their best judgement and experience. Subjectivity is often a large component of such estimations, as are the experience and objectives of the evaluator.

The authors provide an alternative by developing and leveraging an augmented artificial-intelligence (AI), data-driven approach. Geological and engineering features are considered as well as operational conditions to bring domain knowledge into the analysis, resulting in more-accurate and -reliable predictions. In addition, the method is ex­tended to conduct DCA probabilistically using quantile regression techniques.


Deterministic DCA. The work flow is built upon six steps.

Step 1—Data Loading and Data-Quality Checking. Data is loaded into the platform in a predefined format required for advanced analytics, machine learning, and optimization purposes. The input data (including settings and parameters) for technology involve production data and physical and operational information of all wells.

Step 2—Machine Learning and Smart Clustering Analysis. Clustering analysis is performed using machine-learning and pattern-recognition techniques to partition the data set into internally homogeneous and externally distinct groups. These clusters are then used to generate type curves for each group (cluster) of wells to apply to the wells for which production histories are not lengthy.

Step 3—Event Detection. The next step is to identify the last event from which to apply the decline curve. The event-detection algorithm automatically (without human interference) detects any major changes in production-decline behaviors and flags them as a production event. The detection of such events is crucial in DCA because the decline curves fitted before such a break point will not be applicable to the post-break-point period. The idea behind the event-detection methodology is that, if two linear trends—one within a certain interval before the break point (event) and another within a certain interval after the break point—are fit, then the difference in trend estimate on the break point from the two linear trends would be significant if there were to be a sharp change in the production behavior.

Step 4—Cross Validation With Real-Life Events. Because the ­changes in production are correlated highly with real-time events such as perforation changes, workovers, compression, operating hours, and artificial lift, a cross validation with the operating conditions (if available) is performed to ensure robustness and accuracy.

Step 5—Curve Fitting. The engine computes eventually a decline curve for all events using advanced optimization (minimization, in this case) algorithms for nonlinear least-square problems and selects the best fit for the data while considering the physical changes in the field. This not only leads to more reliable predictions but also provides insight to the differential estimated ultimate recoveries (EURs) after the corresponding reservoir or operating condition changes.

Step 6—Type-Curve Generation. Once all wells are analyzed and their declines are calculated, the wells in each cluster are aggregated to generate type curves for their respective cluster. These cluster-based type curves are then applied to the wells in each respective cluster with short production histories in order to conduct meaningful DCA.

Probabilistic DCA. Though automated event detections—augmented with real-life events, robust curve-fitting optimization algorithms, and clustering/pattern recognition techniques—can be used to perform DCA systematically and automatically, the unbiased aspect of the process is not fully assured because the decline is heavily dependent on the choice of the curve-fitting objective function. Depending on the importance of a subset of data points, an infinite number of decline curves can be obtained, with some being pessimistic, some close to the median, and some optimistic. This issue can be alleviated by choosing to minimize the residual sum of squares. However, one could argue that some data points are more informative than others.

Therefore, the method is extended to a probabilistic approach using quantile regression. This solution allows obtaining different confidence intervals for decline curves, thus providing a confidence band on the declines for individual wells and for the entire field.

Quantile regression is a method that estimates the conditional quantile or percentile of the response variable on the basis of data and can help obtain a more-comprehensive understanding of the relationship between variables of interest. The authors use quantile regression to simulate the effect of uncertainty in a production forecast. The authors’ approach with regard to both well-level and field-level DCA is discussed in detail in the complete paper.

Case Study: A Giant Sandstone Reservoir

The complete paper presents a case study in which the authors’ approach was applied to a giant mature sandstone reservoir in the Middle East. The field has nearly 70 years of production and more than 2,000 wells.

Traditional DCA of this field can take several weeks at least if it is performed on a well-by-well basis. The work can be sped up if it is split among multiple engineers, but then different opinions will exist regarding what events are considered and where to start the decline. The described method dramatically decreased the amount of time required to analyze such a large data set and ­removed any possible engineer biases. In a matter of minutes, the tool integrated all production data and real-life information throughout the entire history and provided a comprehensive analysis at the well and field levels. Overall, for this case study, the spread between the low and the high decline rate has varied from approximately 1 to 10%, depending on the amount of historical production.

Fig. 1 shows two metrics derived from the automated DCA engine used by the authors, which helps to understand the reservoir situation better. For instance, the left plot shows the average estimated ultimate recovery per well by the year the well commenced production. The figure shows that the average EUR per well has been decreasing since the early stages of the field. The well-by-well EUR per spud year (right plot) indicates the same trend of decreasing EUR over time. These metrics highlight some deficiencies in pressure depletion over time.

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

Another significant benefit of access to well-by-well decline rates for such a large field is that different groupings can be analyzed, providing insight into the performance of different wells of interest. For example, groupings may be created in terms of well type (vertical, deviated, or horizontal), artificial lift, geographical area, or pads. For this reservoir, the different well types exhibit very similar decline rates (within 1%), yet, not surprisingly, horizontals have the lowest declines. Deviated wells have the highest decline rates, possibly the result of sidetracks arising to counter drilling problems. Most of the wells in the field are vertical, thus representing the highest cumulative oil production among the groups.

A key feature of the authors’ DCA technology is its machine-learning solution that allows users to partition a data set into groups that are homogeneous internally and distinct externally. The classification is carried out on the basis of a measure of similarity or dissimilarity within and between the groups. Different cluster features may be selected or can be identified automatically. On the basis of those features, the DCA engine clusters the wells and uses them for enhanced type curving.

The results from the DCA allowed quantification of the remaining reserves by reservoir with much higher accuracy. They also highlighted multiple potential areas for improvement to increase the ­ultimate recovery of the reservoir.


The main features of the technology ­include the following:

  • Fully automated DCA solution that allows focus on result interpretation and analysis
  • Advanced event-detection algorithm with a cross-validating framework to compare automatically detected events against real-life events
  • Flexible machine-learning-based solution for precise type curving in wells with limited production history
  • Uncertainty quantification framework based on quantile regression providing a confidence interval range for the decline estimation
  • Scalable algorithm to generate field, group, and cluster baseline declines

This article, written by JPT Technology Editor Chris Carpenter, contains highlights of paper SPE 197142, “Artificial-Intelligence-Based, Automated Decline-Curve Analysis for Reservoir Performance Management: A Giant Sandstone Reservoir Case Study,” by Amir Kianinejad, Rami Kansao, and Agustin Maqui, Quantum Reservoir Impact, et al., prepared for the 2019 Abu Dhabi International Petroleum Exhibition and Conference, Abu Dhabi, 11–14 November. The paper has not been peer reviewed.