The major challenge in hydraulic modeling is that achieving a realistic representation of wellbore conditions is difficult in a mathematical model. This paper presents a new approach in monitoring the hydraulic system and in the recognition of well‑control events at an early stage such that proper counteractions can be initiated before any damage occurs. A hydraulic real-time monitor based on sensor data has been developed, using artificial neural networks to compute, recognize, and predict abnormal events in the wellbore.
Introduction
Precise knowledge of hydraulic pressure while drilling is crucial for the success of the entire drilling process. The pump must operate within an accurate pressure range, providing the desired flow rate necessary for hole cleaning. The operating pressure is thereby concurrently limited by pipe and equipment constraints. Pore pressure, formation-fracture gradients, and equivalent circulating density (ECD) introduce additional complexity to the hydraulic system. However, when taking the complete drilling system under consideration, one must remember that severe challenges arise in representing the reality because of the restricted knowledge of several operational parameters at any point of time.
To improve the quality and reduce the influence of the detailed input parameters, the deterministic approach is improved by introducing numerical methods that add significantly to the performance of the system. The hybrid simulator is then in fact able to handle the uncertainties regarding the input data. Furthermore, a major step toward gaining optimal results is achieved by the implementation of automatic-drilling-operation recognition systems. These enhance the output of the numerical models and provide the means for operation-based monitoring and performance analysis.
Hydraulic Models
Deterministic Model. The deterministic approaches are based on classical rheological models incorporating mud properties, borehole geometry, and casing and drillstring configuration, as well as detailed information regarding the well path. The developed model begins with a wellbore-geometry-design routine. Because the wellbore is a combination of several sections, with a complex drillstring tripping in and out, an exact representation of the actual wellbore geometry is of critical importance for the forthcoming calculations.
Once the geometry is determined, the simulator steps into a procedure for representing the drilling-mud rheological properties. It is well-known that drilling muds show non-Newtonian behavior, meaning that their viscosity is not only temperature- and pressure-dependent, but is also a function of the shear rate. Rheological models are mathematical models that describe mud viscous properties in a shear-rate/shear-stress relation. A common practice in the oil industry today is to apply the simplified Bingham rheological model, which fits drilling data quite well. However, the more realistic power-law and Herschel-Bulkley models provide even better solutions.
A crucial part of every drillstring is the bottomhole assembly (BHA). Depending on the type of operations performed, the BHA configurations may include a combination of various tools and tubulars with different lengths, inner and outer diameters, and functions. The main part of the BHA, including the heavyweight drillpipe and drill collar, is typically loaded manually into the system by the user. For all remaining complex tools, a fixed differential-pressure value is added to the total simulated pressure drop. Mud motors are handled separately because their pressure drop is torque-dependent. Manufacturer performance curves are used to handle these nonlinear relationships. Bit and surface pressure losses are also accounted for with the latest American Petroleum Institute guidelines. Finally, the overall pressure drop in the system is calculated.
Although the deterministic approach works well when all input parameters of the model are precisely defined, in real-world applications, a lack of accuracy is caused by several uncertainties. Besides these shortcomings, the borehole diameter in open holes usually cannot be defined exactly and may vary over time. To tackle those usually strong nonlinear challenges, the deterministic models have been extended by data-driven models—namely, neural networks—to a hybrid model. On the basis of real-time sensor measurements, those extended models were trained for simulating the standpipe pressure on both single-well and crosswell scenarios.
Hybrid Model. The biggest disadvantage of every analytical drilling-hydraulics approach is the lack of data throughout the simulation. As described previously, precise parameter selection, together with regular system calibration with respect to real standpipe-pressure readings, is required to ensure accurate simulation results. However, most of the data needed are not available for the user in a timely manner. Even if exact data-gathering procedures are defined, the quality is low and timing is inadequate. To compensate for this deficiency, numerical methods are introduced to the model. Neural networks are software algorithms that are meant to simulate the human brain. Currently, their application is spread across a wide range of industries such as the medical industry, weather prediction, and image and voice recognition. Neural-network techniques are useful when no concrete rules are available and when a heuristic solution is less demanding than a model-based approach.
There are different training algorithms involved with neural networks. The selection of the training data set and the configuration of the network is a crucial part of the process. A small and simple network configuration can result in a lack of precision, while heavy and complex networks tend to have oversizing issues. For each particular application, a lengthy trial-and-error method is normally required.
Automated Operations Recognition (AOR)
Efficiency has always been of great importance in the drilling industry. With the increasing complexity and the rising costs of drilling operations, performance plays a major role in the economic and technical success of every oilwell project. Because many of the operations performed are repetitive, one option is to measure performance with stopwatches. However, the poor results and the resources involved drove the search for a better solution.
The patented AOR system used in this work uses high-sampling-rate rig-sensor data (between 1 and 0.1 Hz) and automatically detects and measures the time taken for the most common drilling operations. Whether drilling in rotary or sliding mode, reaming up and down, or washing, benchmarking becomes possible because different comparisons to appropriate batch wells can be made easily.
The process begins with a regimen that checks the provided data and ensures that poor entries and data gaps will not influence the following operation-recognition steps. As soon as the filtering process is concluded, the AOR procedure takes over and detects, and calculates the time for, each operation.
Implementing AOR Into the Hydraulic Model: Intelligent Hybrid System
Neural-network analyses depend primarily on the quality, rather than the quantity, of data. One should ensure that the selected training data sets correlate precisely with the output. However, because drilling operations differ from one well to another, the deviations in the input parameters, such as flow rates, rotary speeds, and torque, may foul the simulation results. To find an improved result, a decision was made to include the operation-recognition outcome as a part of the training set. Because the automatic event-detection system provides accurate results in real time, its implementation in the hydraulic-monitoring system added significantly to the accuracy and quality of the concept.
Once implemented in the model, the AOR system offers several additional capabilities. Hydraulic monitoring also may be performed per specific operation, meaning that one can focus on specific activities such as drilling, wellbore conditioning, or anything else of interest to the driller.
Simulation and Results
During the testing stages of the concept, a number of training and simulation runs were performed with a wide variety of rig-sensor data. Different rig configurations, setups, and conditions were tested. For a summary of the land-rig scenario, please see the complete paper.
Offshore-Rig Scenario. The focus of this work was deliberately set to offshore rigs, where the equipment, well construction, and conditions are much more challenging. Unlike the easier-to-handle land-rig wells, the complex set of variables in offshore wells requires additional model tuneup and adjusting. Besides the specific neural-network configuration, the system proved to be highly dependent on the data quality and the AOR-system results.
The given example includes two drilling runs performed during an 8.50-in. horizontal section. The first run is used for the initial network training. 80% of the data are selected for training. The remaining 20% are left for testing and validation (10% each). As the training procedure is completed, the created network can be used for simulations and standpipe-pressure modeling on suitable well scenarios.
Fig. 7 of the complete paper shows that though only seven data channels were used and the simulation was performed without any manual data entry, the error lies in acceptable ranges. Moreover, the simulated pressure losses clearly follow the trend of the measured standpipe pressure. This fact confirms the assumption that an accurately designed intelligent hybrid model is capable of handling complex wellbore-hydraulics issues with sufficient accuracy.
Operating Window
After the simulator has been configured and calibrated, the results are presented in a simplified and clear interface to the user. The minimum flow rate required for efficient cuttings transport, the maximum flow rate before fracturing the formation, ECD, pump efficiency and limitations, and the simulated theoretical pressure drop are all plotted on an operating window colored green in Fig. 1.
Finally, a library of simulated abnormal-event scenarios was developed, which, in combination with the AOR system and the simulated pressure losses, can be used in real time for trouble identification and well-control-event detection. Once the real-time measurements approach a predefined value, an alarm indicator will notify the driller of a possible unplanned event. Additional capabilities may include pump-startup sequences and various operational modes.
Conclusion
Because the real-time data availability and data quality differ for each customer, the current stage tests were made with only seven data channels for training (bit and hole measured depth, block position, flow in, hookload, rotational speed, and torque), which were generally available, accessible, and usually had sufficient quality. Furthermore, the operation-recognition outputs were implemented, resulting in improved data training and simulation algorithms.
The authors can conclude that using only the simplest possible setup (seven data channels plus the AOR system), the result is that more than 95% of the theoretical standpipe-pressure values reside within 10 bar of the real measured data. Furthermore, approximately 60% of the calculated data points lie within 5 bar of the RT data. The statistics are based only on drilling data, because drilling represents the main field of interest here. In cases where wellbore-conditioning activities take place, the performed pretests showed that they can be simulated at least as well as the drilling states.
On the basis of the fact that no BHA-detail, mud, survey, or manual input was used, and the fact that the model can be applied immediately on any on- or offshore well, the level of accuracy (together with the potential for improvements after introducing features and additional data channels) was found to be sufficient and highly promising.
This article, written by JPT Technology Editor Chris Carpenter, contains highlights of paper OTC 24803, “Hydraulic Monitoring and Well-Control-Event Detection by Use of Model-Based Analysis,” by D. Todorov and G. Thonhauser, Thonhauser Data Engineering, prepared for the 2014 Offshore Technology Conference Asia, Kuala Lumpur, 25–28 March. The paper has not been peer reviewed. Copyright 2014 Offshore Technology Conference. Reproduced by permission.