Automated Well Top Picking and Reservoir Property Analysis of the Belly River Formation of the Western Canada Sedimentary Basin
Baosen Zhang*1, Tianrui Ye1, Yitian Xiao1, Dongmei Li2, Guoping Wang2, Cong Su2, Tongyun Yao3,
1. Petroleum Exploration and Production Research Institute, SINOPEC, 2. International Petroleum Exploration and Production Corporation, SINOPEC, 3. ESSCA Group
Copyright 2022, Unconventional Resources Technology Conference (URTeC) DOI 10.15530/urtec-2022-3719133
This paper was prepared for presentation at the Unconventional Resources Technology Conference held in Houston, Texas, USA, 20-22 June 2022.
The URTeC Technical Program Committee accepted this presentation on the basis of information contained in an abstract submitted by the author(s). The contents of this paper have not been reviewed by URTeC and URTeC does not warrant the accuracy, reliability, or timeliness of any information herein. All information is the responsibility of, and, is subject to corrections by the author(s). Any person or entity that relies on any information obtained from this paper does so at their own risk. The information herein does not necessarily reflect any position of URTeC. Any reproduction, distribution, or storage of any part of this paper by anyone other than the author without the written consent of URTeC is prohibited.
Abstract
Accurate well top picking and reservoir property prediction plays crucial role for petroleum exploration and production. The task is traditionally done manually by geologists, and can be inconsistent and time-consuming. In addition, the work is difficult to normalize and standardize due to human bias. For unconventional resource plays with tens of thousands of wells, constructing geological models and algorithms can be a daunting task. Machine learning is an emerging technology that has been increasingly adopted in the energy industry. It can provide automated and accurate well top picking and reservoir property analysis.
This paper utilizes a case study in the Belly River Formation (BRF) of Western Canada Sedimentary Basin (WCSB) to discuss capabilities EnABLEd by automated well top picking and reservoir property analysis. First, 70993 wells with GR curve covering ~100000 km2of WCSB were filtered. Among them, 32510 coring wells help to determine boundaries of BRF. Second, several tops were manually picked as seeds for automated picking using Subsequent Dynamic Time Warping approach. After quality control and log normalization, automatic picks were promoted into new seeds for subsequent picking until all pickings were done. Finally, the distribution of the BRF were defined, and combining with logging curves, the variation law of reservoir properties (porosity, permeability, saturation, etc.) was analyzed.
Automated well top picking algorithm natively handles log normalization issues and picks. It completed ~70000s wells top picks in about 100 hours on cross section and map view, which may take over 1000 hours using traditional manual picking methods. Moreover, after automated well top picking, reservoir properties can be predicted as a “one-mouse-click” exercise. What need to do is to ascertain the acquired reservoir properties according to the production practice and to determine the algorithms and formulas according to the regional geological features. This workflow greatly improves efficiencies of the comprehensive reservoir evaluation and reservoir geological modeling of the WCSB by orders of magnitude.
Subsequently, combining automated well top picking and reservoir property analysis results and real-time data of oilfield production, the exploration and production sweet spot prediction of the BRF of the WCSB can be done. In conclusion, this efficient approach based on machine learning has been successfully applied to the potential assessment of petroleum resources in the BRF. The assessment results were used for petroleum reservoir exploration and production, oilfield development plan design, and portfolio management and optimization. Application of the method requires cooperation across different disciplines—data science and earth science. The interdisciplinary nature provides accurate prediction and design optimization for unconventional resources exploration and production.
Introduction
Since the 21st century, large-scale computing, big data and deep learning have triggered the Third AI Boom (Brynjolfsson and Mitchell, 2017). Recently, AI has also been widely used in petroleum exploration and production. Operators cooperate with academic institutes, IT companies and vendors to carry out AI application research, which is developing rapidly in the direction of digitalization, integration, visualization and artificial intelligence application (Li et al., 2020).
Accurate well top picking and reservoir property prediction plays crucial role for petroleum exploration and production. The task is traditionally done manually by geologists and can be inconsistent and time-consuming. In addition, the work is difficult to normalize and standardize due to human bias. For unconventional resource plays with tens of thousands of wells, constructing geological models and algorithms can be a daunting task. Machine learning is an emerging technology that has been increasingly adopted in the energy industry. It can provide automated and accurate well top picking and reservoir property analysis.
Previous attempts have been made to pick geologic well tops automatically using expert systems (Olea, 2003), neural networks (Luthi, 2001), and dynamic programming (Lineman et al., 1987; Inazaki, 1994; Steven et al., 2004; Fang, 2009). Although these previous efforts have been helpful in defining the problems and establishing the building blocks to solve well-log correlation automatically, owing to the nature of seismic data, they have clearly been observed to be much less successful than seismic picking algorithms. Comparing to seismic traces, well logs are more widely spaced (on the order of hundreds to thousands of meters), have inconsistent depth ranges with possible gaps, and may be from highly non-vertical well bores. As a variant of the Dynamic Time Warping (DTW) algorithm, Subsequence Dynamic Time Warping (SDTW) was introduced by Grant et al. (2018) to perform the relevant curve alignments. This technology and workflow uses the power of the modern computer and novel machine learning techniques to capture and model well-log patterns for correlating geologic events across thousands of wells. Using one or more well logs as source wells, a signature ‘thumbprint’ segment is correlated over many target wells to find the optimal stratigraphic intervals for well pick estimation.
This paper utilizes a case study in the BRF of the WCSB to discuss the capabilities enabled by automated well top picking and reservoir property analysis. This study is implemented to support the project of “Western Canada Sedimentary Basin– Edmonton/Belly River Potential Analysis” from SINOPEC International Petroleum Exploration and Production Corporation. First, an overview of the geological setting of the BRF of the WCSB was provided. Second, the methodology of the automated well top picking, including the theoretical basis and workflow, was explained in detail. Third, reservoir property analysis is applied to the BRF of the WCSB based on the automated well top picking results, empirical formulas or machine learning algorithms for property inference, and evaluations by geologists and engineers of petroleum exploration and production. Finally, to further deepen this research results, in the future the production sweet spot prediction model will be established based on production data and reservoir property analysis results, and the intelligent prediction of petroleum favorable areas in the BRF of the WCSB will be completed (Figure 1).
Figure 1. Flowchart summarizing the workflow of automated well top picking and reservoir property analysis of the BRF of the WCSB in this paper. It begins with automated target well top picking, followed by reservoir property analysis. Then the production sweet spot prediction is a future plan based on the above two works. Finally, a scientific and efficient data analysis and machine learning workflow will be developed.
Geological setting
The main deformation events recorded in the WCSB took place during two orogenic periods (Figure 2): Nevadan Orogeny (Late Jurassic to Early Cretaceous) and Laramide Orogeny (Late Cretaceous to Paleocene). Before the Nevadan Orogeny, the WCSB was a stable Craton basin deposited in the passive continental margin. As its provenance were mainly the Craton interior from the east, it is dominated by the shallow sea facies. After the Nevadan Orogeny, the WCSB was converted to a foreland basin deposited in the active continental margin. The input source of the littoral and fluvial facies has changed to the fold and thrust belt, which was formed by the compressive deformation fromJurassic to Paleocene, resulting in eastward transportation of the sedimentary units towards the WCSB.
Figure 2. Geological evolution diagram of the WCSB from Paleozoic to Cenozoic. The stratigraphic column is modified from the SINOPEC International Petroleum Exploration and Production Corporation.
Refining to the target layer, the BRF is a Late Cretaceous Campanian stage about 72-84 ma continental sandstone deposits. As the early Cretaceous was dominated by the marine deposits, the late Cretaceous BRF was sometimes affected by seaway intrusion and developed a small amount of mud shale deposits. The BRF is mainly composed of three sandstone-shale sub-cycles. The bottom boundary of the BRF, overlying the Lea Park Formation, is marked by the end of continuous marine shale deposition and the emergence of marine-continental transitional sandstone deposition. It remains extremely consistent in the whole WCSB scale, and the characteristics of its well logging curves are very distinctive (Figure 3). The top boundary of the BRF, underlying the Edmonton Formation, marks the emergence of large-scale transgression, the weakening of sandstone deposition, and the increasing of shale deposition. However, its consistency at the basin scale is poor, and its well logging curves’ characteristics are not obvious (Figure 3). Since the bottom boundary is easy to judge, most of effort were put into developing criteria for ascertaining the location of the top boundary. Finally, the top boundary was defined as the position where the maximum flood surface first appears, signed by the high natural gamma (GR) appearing and the amplitude of the resistivity (RD) decreasing (Figure 3).
Figure 3. The well logging curves characteristics of the BRF and its underlying and overlying strata, and its comprehensive judgment position of top and bottom boundaries. Blue and red triangles represent the sub-cycles of transgression and regression.
Methods
Using Accumap for data collection, 171814 well heads, 75301 wells with curves and 36818 wells with core data were combed out (Figure 4). Then,70993 wells drilled into the BRF with GR curves covering ~100000 km2 of WCSB were filtered. Among the filtered wells, 32510 coring wells help to determine the boundaries of BRF.
After data collection, SDTW algorithm was used for the automated well top picking. Figure 5 below shows two different hypothetical data series that might represent two well logs from adjacent wells. The DTW algorithm considers all possible stretch, squeeze, and shift combinations to optimally align the corresponding peaks and troughs along a minimal cost path (Grant et al., 2018). This approach is needed to capture laterally varying geologic changes from well to well as stratigraphic thinning and thickening occurs. However, the DTW does not handle situations where the start and end points represent different times or, in well logging curves, depths. To handle this limitation, Müller (2007) adjusted the DTW algorithm by, instead of aligning the sequences globally, finding a subsequence within the longer sequence that optimally fits the shorter sequence. The variation of the DTW algorithm is called Subsequence Dynamic Time Warping (SDTW), which has many applications in database querying.Database querying, intrinsically similar to well logging curves correlation, entails identifying the fragment in the dataset that is most likely the query by matching a smaller pattern representing the query into a much larger data sequences in the database (Grant et al., 2018).
Figure 4. Well data compilation and their tectonic locations of the WCSB.