Application of Data Science and Machine Learning for Well Completion Optimization

Abstract

In today's data-driven economy, operators that integrate vast stores of fundamental reservoir and production data with the highperformance predictive analytics solutions can emerge as winners in the contest of maximizing estimated ultimate recovery (EUR). The scope of this study is to demonstrate a new workflow coupling earth sciences with data analytics to operationalize well completion optimization. The workflow aims to build a robust predictive model that allows users to perform sensitivity analysis on completion designs within a few hours. Current workflows for well completion and production optimization in unconventional reservoirs require extensive earth modeling, fracture simulation, and production simulations. With considerable effort and wide scale of sensitivity, studies could enable optimized well completion design parameters such as optimal cluster spacing, optimal proppant loading, optimal well spacing, etc. Yet, today, less than 5% of the wells fractured in North America are designed using advanced simulation due to the required level of data, skillset, and long computing times. Breaking these limitations through parallel fracture and reservoir simulations in the cloud and combining such simulation with data analytics and artificial intelligence algorithms helped in the development of a powerful solution that creates models for fast, yet effective, completion design. The approach was executed on Eagle Ford wells as a case study in 2016. Over 2000 data points were collected with completion sensitivity performed on a multithreaded cluster environment on these wells. Advanced machine learning and data mining algorithms of data analytics such as random forest, gradient boost, linear regression, etc. were applied on the data points to create a proxy model for the fracturing and numerical production simulator. With the gradient boost technique, over 90% accuracy was achieved between the proxy model and the actual results. Hence, the proxy model could predict the wellbore productivity accurately for any given change in completion design. The operators now had a much simpler model, which served as a plug-and-play tool for the completion engineers to evaluate the impact of changes in completion parameters on the future well performance and making fast-tracked economic decisions almost in real time. The approach can be replicated for varying geological and geomechanical properties as operations move from pad to pad. Although the need for heavy computing resource, simulation skillset, and long run times was eliminated with this new approach, regular QA/QC of the model through manual simulations makes the process more robust and reliable. The methodology provides an integrated approach to bridge the traditional reservoir understanding and simulation approach to the new big data approach to create proxies, which allows operators to make quicker decisions for completion optimization. The technique presented in this paper can be extended for other domains of wellsite operations such as well drilling, artificial lift, etc. and help operators evaluate the most economical scenario in close to real time.

Machine Learning Tools for Fossil and Geothermal Energy Production and Carbon Geo-sequestration—a Step Towards Energy Digitization and Geoscientific Digitalization

“…Current workflows for well completion and production optimization in unconventional hydrocarbon reservoirs require extensive earth modeling, fracture stimulation, and production simulations, which are costly and time-consuming. Pankaj et al [14] proposed a data-driven approach on Eagle Ford wells as a proxy model for optimizing the completion design. This was achieved using supervised learning.…”

“…Accuracies of these Fig. 2 Comparison of prediction accuracies using gradient boosting method when predicting production over 1 month (B1), 3 months (B2), 12 months (B12), and 5 years (Cum_5), as performed by Pankaj et al [14]. X-axis represents error tolerance, and Y-axis represents percentage of samples within the error tolerance.…”

Machine learning on field data for hydraulic fracturing design optimization

“…Source Cumulative oil production 6/18 month just after the job [24] 12 months cumulative oil production [25] Average monthly oil production after the job [19] NPV [26] Comparison to modelling [28] Delta of averaged Q oil [29] Pikes in liquid production for 1, 3 and 12 months [30] Break even point (job cost equal to total revenue after the job) [32] • In [26], a procedure was presented to optimize the fracture treatment parameters such as fracture length, volume of proppant and fluids, pump rates, etc. Cost sensitivity study upon well and fracture parameters vs NPV as a maximization criteria is used.…”

“…• Q pikes approach is presented by implementing B1, B2 and B3 statistical moving average for one, three and twelve-month best production results consequently in [30]. The simulation is done over 2000 dimension dataset to reap the benefit from proxy modeling treatment.…”

“…Metrics Source Cumulative oil production 6/18 month just after the job [24] 12 months cumulative oil production [25] Average monthly oil production after the job [19] NPV [26] Comparison to modelling [28] Delta of averaged Q oil [29] Pikes in liquid production for 1, 3 and 12 months [30] Break even point (job cost equal to total revenue after the job) [32] Table 1: Success metrics of HF job…”