|Theme||Visible||Selectable||Appearance||Zoom Range (now: 0)|
Abstract Drilling a directional well becomes an essential process in the oil and gas industry to ensure better reservoir exposure and less wellbore collision risk. In the high-volume drilling market, cost-effective mud motors are dominant. The motor is capable of delivering the desired well curvature by switching between rotating and sliding operations. Therefore, to follow a predefined well trajectory, it is a critical mission to determine the optimal operation control sequence of the motor. In this paper, a method of training an automatic agent for motor directional drilling using the deep reinforcement learning approach is proposed. In designing the method, motor-based directional drilling is framed into the reinforcement learning with an automatic drilling system, also known as an agent, interacting with an environment (i.e., formations, wellbore geometry, equipment) through choices of controls in a sequence. The agent perceives the states such as inclination, MD, TVD at survey points and the planned trajectories from the environment, and then decides the best action of sliding or rotating to achieve the maximum total rewards. The environment is affected by the agent's actions and returns corresponding rewards to the agent. The rewards can be positive (such as drilling to target) or negative (such as offset distance to the planned trajectory, cost of drilling, and action switching). To train our agent, currently, a drilling simulator in a simulated environment is created with layered earth model and BHA directional responses in layers. Other attributes of the drilling system are assumed to be constant and handled automatically by the simulator. The planned trajectory is also provided to the agent while training. The directional-drilling agent is trained for thousands of episodes. As a result, the agent can successfully drill to target in this simulated environment through the decisions of sliding and rotating. The proposed workflow is known as the first automated directional drilling method based on deep reinforcement learning, which makes a sequence of decisions of rotating and sliding actions to follow a planned trajectory.
Zhao, Ying (China University of Petroleum-Beijing) | Sun, Ting (China University of Petroleum-Beijing) | Yang, Jin (China University of Petroleum-Beijing) | Yin, Qishuai (China University of Petroleum-Beijing) | Wei, Hongshu (CNOOC China Limited, Shenzhen Branch) | Liu, Zhengli (CNOOC China Limited, Shenzhen Branch) | Li, Zhong (CNOOC China Limited, Zhanjiang Branch) | Huang, Yi (CNOOC China Limited, Zhanjiang Branch)
Abstract For an oil well, we can determine the working conditions of drilling wells and whether there is an accident by logging data. However, it takes a long time to analyze logging curves by traditional manual work. Therefore, this paper proposes a new method combining logging big data (BD) and machine learning (ML), which can train and study a large number of logging curves and form a data base. Then this data base can be used to judge the working conditions of the well being drilled and whether there is an accident automatically. In the proposed method, firstly, due to the large span of data and different detection methods, there will be some differences in data accuracy, so the data need to be preprocessed to avoid these influences. Secondly, the preprocessed data is input into the artificial intelligence method, and trained by cyclic iterations; meanwhile the rule of judging working conditions is also learned from the data. Thirdly, we get a training set containing all kinds of working conditions and common accidents through artificial intelligence learning method. Finally, the logging data that need to judge is input into the training set after preprocessing, and then we can determine the working conditions and whether there is an accident according to the output of the procedure. For the proposed method, we used the logging data of Bohai Oilfield to verify that a total of 200 logging data including 4 conditions were preprocessed and input to the artificial intelligence method to randomly generate a training set (160 sets) and the test set (40 groups) is output after the training set. The results show that the relative error of the four kinds of working conditions output by the proposed method is less than 5%, and the average calculation time is faster than traditional method, saving a lot of manual processing time. This article introduces the methods of big data and artificial intelligence into the field of drilling, which have been tested by actual data. The time for judging drilling conditions by manpower can be greatly saved, and the common accidents during drilling can be found in time, which show that this method is of great significance.
Cao, Dingzhou (Occidental Petroleum Corporation) | Hender, Don (IPCOS) | Ariabod, Sam (Apex Systems) | James, Chris (Occidental Petroleum Corporation) | Ben, Yuxing (Occidental Petroleum Corporation) | Lee, Micheal (Occidental Petroleum Corporation)
Abstract This paper provides the technical details to develop a real-time deep learning model to detect and estimate the duration of downlinking sequences of Rotary Steerable Systems (RSS) based on a single measurement (standpipe pressure, SPP). Further analytics are derived based on the downlink recognition results together with other real-time log data (ROP, RPM, Torque, etc.) to drive directional drilling efficiency. Real-time RSS downlink recognition is treated as an image segmentation problem. The Deep Learning (DL) models were created using the dynamic U-Net concept and materialized with a pre-trained ResNet-34 as the underlying architecture. Transfer learning was used due to the limited number of training samples (≪ 100 downlinks per onshore well) to help with speed and accuracy. The SPP time series data was segmented based on stand of pipe drilled (one image per stand). This "image" was then fed into the model for downlink recognition. To further increase the accuracy, a second opinion mechanism was applied when the models were tested and deployed into the Real-Time Drilling (RTD) system. Using a dual model approach greatly reduced the number of false positives due to non-downlink pressure fluctuations causing "noise". The patterns of SPP and its rate of change (delta SPP) are quite different. They both have pros and cons for identifying the downlink, thus two independent models were built based on these two signals. The DL model A is trained based on the original SPP signal and the DL model B is trained based on delta SPP. A downlink is confirmed only when both models show positive results. Data of 10 onshore wells (2 rigs) drilled with RSS were segmented (8165 images in total) and labeled. There were 671 images with 795 downlinks and 7980 images without downlink. The five-fold cross-validation technique was used to identify the best model(s). The F1 score of blind test result was .991 (accuracy was ~99.82%, see Table 2). The relative error of duration estimation is 2.49%. The current rig fleet within the RTD system has a mix of drilling tool configurations - RSS and mud motors. To further validate the models’ robustness regarding drilling tools, additional tests were conducted using mud motor wells’ datasets from 21 rigs (25431 images without downlink). There were 3 false negatives from this extended test set, which resulted in a ~99.93% accuracy for the aggregated 31 wells dataset. These results suggest that the models are accurate, reliable and robust. The real-time DL solution presented in this paper enables operators to analyze RSS performance during and between downlinking events. This would allow drilling engineers and rig supervisors to make faster, more reliable data-driven decisions to optimize performance and directional control of the well path.
ABSTRACT Success of directional wells is reliant on accurate identification of formation being drilled, and correction of off-shoots and out of zone trajectories by steering the well back in the correct direction. Geo-steering is the directional control of a well path based on real-time measurements assimilated while drilling and comparing the observed variables to expectations drawn from nearby wells and known formations. Directional drillers and geologists alike must make split-second geo-steering decisions that ultimately affect overall productivity of the well. Providing an accurate, predictive model to cross-check with the inflowing real-time data allows specialists to confidently keep the wellbore in the zone of interest. The present study uses geological and geophysical data from offset wells, to estimate lithological facies for the projected well trajectory. A suite of machine learning and data analytics algorithms, including Long Short-Term Memory (LSTM), are employed to extract the characteristic signatures of different lithofacies, and generate the expected signatures for the well of interest, based on the available logs from nearby wells. Upon detection of deviation from the expected path, corrective measures can be suggested to the operator, allowing for confident geo-steering. 1. INTRODUCTION The advent of downhole drill motors and directional drilling has made it possible to expose several thousand feet of reservoir to the well as potential pay, versus a few feet of vertical section that was achieved with vertical wells. This can significantly increase the contacted reservoir interval and productivity of the well. Success of directional drilling relies on accurate Geosteering, where real time inputs from Measurements While Drilling (MWD) tools are used to determine geolocation and steer the drill bit to keep the well on the desired path, and in the pay zone. The MWD tools are placed above the drill bit. MWD tools are also commonly referred to as Logging While Drilling (LWD) tools, however, it should be noted that MWD is more specifically geared towards the drilling operation aspects and include different measurements (inclination, azimuth, downhole weight on bit, torque, …), while LWD uses logging probes (standard and azimuthal gamma ray, resistivity, density, …). In earlier implementations of MWD/LWD where communications through mud pulsing was limited to a few bits per second, LWD was recorded during the drilling operation and was processed when the drilling assembly was brought to the surface. More recent advances in data communication with the bottomhole assembly has resulted in real-time transmission of MWD/LWD to the surface (Jellison et al. 2003). LWD provides accurate measurements of the formation before it is exposed to excessive drilling fluid and its invasion, or enduring long drilling operation, and are superior to wireline logging data in that regard.
This article is a synopsis of paper SPE 62833, "Barrier-Free Learning for Well Construction," J.A. Trantham, SPE, and J.M. Deagen, Phillips Alaska Inc., originally presented at the 2000 SPE/AAPG Western Regional Meeting, Long Beach, California, 19-23 June.