Multi-task learning for virtual flow metering

This report documents a project to improve subsea flow measurements to allowthe flow rates of individual wells to be known more accurately. Becauseproduction revenues are determined from flow rates, improved flow measurementreduces the financial risk to stakeholders, including producers and the U.S.government. Improved measurement also helps to improve reservoir recovery. Inthis project, commercially available flow models were evaluated for theirsuitability as virtual flow meters, to augment or replace hardware devices forthe determination of flow rates in deep water wells. The goal of this projectwas to develop methods for evaluating flow models, and to report on theefficacy of these models to predict flow rates. Several flow models were usedin the study, which also includes a survey of the available models. 1. Introduction 1.1. What is a Virtual Flow Meter? Virtual flow meters (VFM) are a method for determining the flow rate of a wellby means of modeling the flow. A VFM model may consist of a single well to anentire field of co-mingled wells. The models are fed data from sensorsinstalled at various measurement points (nodes). Typically pressure andtemperature sensors are used, which are available downhole (perhaps at multiplepoints), at the wellhead, and at the surface. With the pressure and temperaturedata at each node provided as an input, the VFM model computes the flow rate ofthe well. Typically, the model is tuned periodically using available flow data. Once tuned, the model is capable of determining the well flow rate with onlythe pressure and temperature nodal data. The flow models employed with VFMsystems may be built from first principles, but this is not the only method forVFM modeling. Flow models may also be derived from statistical analysis ofmeasurements acquired at different nodes along the production system. An important observation is that VFM software is not something that must beinstalled at the same time as the physical multiphase flow meters, as long ascare is taken to install proper sensors to be used at the various nodes. Thefault tolerant nature of the VFM system is an advantage, since a failed "hard"sensor will reduce the number of nodes, but will not cause a complete loss ofinformation. While VFM technology appears to be maturing rapidly, it is not yet generallyaccepted an equivalent replacement for good physical metering. VFM models areemployed as a back up to a physical meter to be used in case the primarymetering system is unable to function, or as an augmentation of physicalmetering to reduce measurement uncertainty. Some of the commercial VFMs use the technique of Nodal Analysis for the flowrate prediction. This method utilizes available transmitters (pressure andtemperature) and a mathematical model of the complete flow path from bottomhole to the topside facility. This method requires considerable input data, including well and pipeline profiles, the diameters of all components ofsections, sensor locations, fluid properties, choke geometry, etc.

Steady-state process models are common in virtual flow meter applications due to low computational complexity, and low model development and maintenance cost. Nevertheless, the prediction performance of steady-state models typically degrades with time due to the inherent nonstationarity of the underlying process being modeled. Few studies have investigated how learning methods can be applied to sustain the prediction accuracy of steady-state virtual flow meters. This paper explores passive learning, where the model is frequently calibrated to new data, as a way to address nonstationarity and improve long-term performance. An advantage with passive learning is that it is compatible with models used in the industry. Two passive learning methods, periodic batch learning and online learning, are applied with varying calibration frequency to train virtual flow meters. Six different model types, ranging from data-driven to first-principles, are trained on historical production data from 10 petroleum wells. The results are two-fold: first, in the presence of frequently arriving measurements, frequent model updating sustains an excellent prediction performance over time; second, in the presence of intermittent and infrequently arriving measurements, frequent updating in addition to the utilization of expert knowledge is essential to increase the performance accuracy. The investigation may be of interest to experts developing soft-sensors for nonstationary processes, such as virtual flow meters.

Occupancy sensor-enabled demand control ventilation using virtual outdoor air flow meters in air handling units without an outdoor air flow meter

A multiphase flowmeter capable of adapting to varying well conditions and providing reservoir transient data is an essential tool for reliable production planning and optimization. This is the main driver for the development of a multiphase virtual flow metering (VFM) module, which is intended to be an add-on to the traditional production management system, (PMS), and as a value addition for the operator. Current physical flow metering systems have certain limitations such as a need for extensive calibration, and high capital and maintenance costs. Also, simultaneous monitoring of every producing well can be cost-prohibitive; thus, making it difficult to determine the field production efficiently and accurately. Having limited real-time field data adds a high degree of unreliability to any simulator predictions. The PMS integrated multiphase VFM described in this paper has been designed to specifically address some of these concerns in a cost-effective manner. This VFM system operates on simulated data sets as well as live field data, which are obtained at a much higher frequency to capture even short-duration transients in the field. An additional advantage of this system is that it can be deployed individually for each well, which facilitates single-well production metering. The VFM module, based on the dynamic multiphase flow simulator, has been developed by one of the industry leaders in simulating transient behavior of flow from a reservoir through the pipeline system. Virtual results such as pressures, temperatures, and multiphase flow rates from a digital-twin of the field are used as base results to predict the field flow rates. These predicted flow rates are compared with real-time data from instruments along the well tubing at the bottom-hole and wellhead. The VFM module processes deviations between predicted and real-time data, adjusts for factors such as instrument sensitivities and field data quality, and corrects the virtual flow rates even further towards more accurate predictions. To validate and confirm this, the model also outputs the confidence rating associated with the improved predictions. These predicted results and associated confidence levels are the final outputs from the VFM system. When integrated with the PMS, this VFM module can also configure alerts to the user in the case of any sudden risks or unwanted transients arising from the reservoir such as water breakthrough in a gas well. Upgrading the PMS with the new VFM module will empower the operator with reliable and accurate flow predictions and help overcome the production flow rate uncertainty In addition to removing dependency on field instrumentation, this VFM module will facilitate improved field production forecasting, and can help lead to timely intervention in potential production-loss scenarios caused by unexpected transients.

Virtual flow metering is a technology for indirect estimation of multiphase production rates. The most straightforward approach to obtain virtual measurements is to apply machine learning models, which enable fast but not necessarily accurate prediction of flow rates. In order to overcome any related limitations machine learning methods can be combined with physical conservation laws of mass, momentum and energy. The resulting hybrid virtual flow metering approach strongly relies on the capabilities of underlying physical simulator. In this work various approaches of combining measurements with multiphase flow models are investigated. It is demonstrated that a wellbore flow model has a central role in any type of hybridization scheme which connects the available measurements to the target flow rates. Due to specific requirements of the virtual flow metering applications, this simulator has to be fast and accurate enough to qualitatively describe flow in wellbores and pipelines. In this work, such a simulator is presented, which is based on a one-dimensional formulation for mass and momentum conservation, but contrary to other commercially available options, the current development is based on a drift-flux approach, which is less demanding on the interfacial closure terms and can be more easily tuned with field data. The performance of the simulator has been evaluated using a set of transient benchmark scenarios, where the reasonable match against the reference data has been demonstrated. The simulator is applicable to a wide range of multiphase flow problems of practical interest and uses advanced numerical algorithms to ensure robustness and performance. The closure relationships can be adjusted to improve the performance for a particular well. In addition, a dedicated graphical user interphase enables user-friendly interaction for setting up the computational case and processing the output production data.

This paper presents a performance evaluation of multiphase flow models when using these models on Virtual Flow Meters (VFM). Recently, VFM simulators have been identified as a technology capable of provide metering backup, monitoring for abnormalities of physical multiphase flow meters and also for production well allocation. Nevertheless, there are just a few studies found in the literature analysing the accuracy of different VFM models applied to multiphase flow metering. In the oil and gas industry it is essential to accurately measure the amount of oil, gas and water being produced by individual wells. Prior to the 1980s, single-phase measurements alone were sufficient to meet the industry standards. However, with the advent of deepwater exploration and production, multiple wells comingling to a single production separator is a common practice in petroleum production for offshore systems. Therefore, the use of multiphase flow meters to account for the production of oil, gas and water from individual wells became evident. This investigation uses field data for more than one well to evaluate the performance VFM simulators. The results obtained from this study shows that VFM models can provide accurate predictions for multiphase flow metering. The simulators evaluated in this study could predict flow rates of oil, gas and water within acceptable levels of uncertainty, with an average absolute error of 20%.

The operation of gas-lift wells is associated with a large number of complicating factors. Hydrate formation, non-productive injection, strong interaction through the oil-gathering system combined with the constantly changing state of operation of each well, requires continuous group optimization of such a fund. The aim of the work is to create information system that allows to automatically generate programs for group optimization of the operating modes of gas-lift wells followed by their expertise and monitoring. The developed system consists of the following key components: a module for working with initial data, which is responsible for the continuous acquisition and aggregation and validation of disparate high-frequency data using virtual flow metering methods; the algorithmic part, which is a complex bundle of components of the steady-state reservoir model, gas-liquid lift model (VLP and nodal analysis), oil-gathering system, machine learning algorithms, data analysis and solving problems of conditional group optimization of wells; business logic that combines individual functional elements within a single process chain. A system is presented that consists of a set of modules for solving the problems of high-frequency data validation, automated creation and adaptation of an integrated model, virtual flow metering, determining the presence of complications in well operation by pattern analysis, determining allowable well operation modes, integrated modeling, and generating a group optimization program using machine learning. The business logic combines the above functional components within a single chain, the key components of which are: continuous determination of current production and the state of operation of wells in the fund; formation of downhole measures to optimize their operating modes, considering technological and geological limitations – both well and group. The integrated approach implemented as part of the development of the presented system was successfully tested in the Eastern section of the Orenburg OGCF. The system shows high efficiency of operational intraday optimization of a large volume of the fund, which is confirmed by the quantitative metrics of the results of the work presented in the article. This system is not limited to work only with gas-lift wells, but also takes into account the impact of events on the ESP and the fountain fund.

Historically well production flow surveillance has been attracting attention from reservoir engineers. The accurate estimation of individual well contribution, Has profound impact on reservoir management. Often production allocation suffers from multiple issues such as quality of available production data and production rate estimation in case of inaccurate and uneconomical measurements. These issues can be mitigated if the well has accurate, calibrated Multiphase Flow Meters (MPFMs) and Virtual Flow Meters (VFMs). Reliability of digital meters is primarily dependent on regular well testing to keep the model updated and tuned to latest reservoir characteristics. As an industry wide common practice, well tests are performed in a timely manner to ensure to allocate production accurately to each well. But performing solo well tests for subsea wells is not trivial as it may seem. Differed production due to downtime associated with flowing subsea wells to a test separator and inability of mature wells to flow to the test separator result in additional challenges. Under such situations, commingled well tests are performed to estimate well rates. Individual well rates can be estimated using test by difference (TBD) method. The apparent simplicity of the TBD method often results in incorrect estimation of allocation rates as the test conditions and production conditions differs significantly. As a result, the credibility of the forecast and the ability to optimize production are compromised. In this paper a novel application called Commingled Well Test Analysis (CWTA) is reviewed which addresses these aforementioned issues. This integrated application comprises of data loading interface, data visualization interface, core optimization engine and an interactive user interface. To demonstrate the capability, three wells from a deep water conventional wells are analyzed using commingled well tests and well performance parameters (PI, WCUT and GOR) are obtained. Based on the tuned parameters, well rates are predicted and compared against rates obtained from sales meter and MPFM meters. As per the comparison, well allocation rates are validated and health of MPFM meters are assessed. As an additional benefit, proper designing of commingled well tests enable in 25-30% time savings of well testing.

Modern challenges in reservoir management have recently faced new opportunities in production control and optimization strategies. These strategies in turn rely on the availability of monitoring equipment, which is used to obtain production rates in real-time with sufficient accuracy. In particular, a multiphase flow meter is a device for measuring the individual rates of oil, gas and water from a well in real-time without separating fluid phases. Currently, there are several technologies available on the market but multiphase flow meters generally incapable to handle all ranges of operating conditions with satisfactory accuracy in addition to being expensive to maintain. Virtual Flow Metering (VFM) is a mathematical technique for the indirect estimation of oil, gas and water flowrates produced from a well. This method uses more readily available data from conventional sensors, such as downhole pressure and temperature gauges, and calculates the multiphase rates by combining physical multiphase models, various measurement data and an optimization algorithm. In this work, a brief overview of the virtual metering methods is presented, which is followed by the application of several advanced machine-learning techniques for a specific case of multiphase production monitoring in a highly dynamic wellbore. The predictive capabilities of different types of machine learning instruments are explored using a model simulated production data. Also, the effect of measurement noise on the quality of estimates is considered. The presented results demonstrate that the data-driven methods are very capable to predict multiphase flow rates with sufficient accuracy and can be considered as a back-up solution for a conventional multiphase meter.

A Machine Learning Approach for Virtual Flow Metering and Forecasting

Due to need for optimal management of a large wells stock with a relatively limited amount of conditioned data, the need for a focus rapid detection of implicit complications in the work (mechanical wear of the working bodies of the pump) and an increase in overhaul period of submersible equipment increases. Real-time monitoring of all wells in the debit is limited due to infrastructure problems and high costs of measuring activities. Despite this, the possibility of such monitoring is not excluded due to the availability of field information that correlates with the well flow rate and the mode of its operation. The article presents an algorithm that, based on telemetry ESP to estimate the parameters that affect the performance of the well and submersible equipment, such as the coefficient of degradation of the outlet characteristics of the pump, the actual efficiency (coefficient of performance) of ESP and the thickness of the deposition of paraffin on the inner walls of the tubing. When processing field data, the parameters correlating with the well flow rate were revealed, which allowed to build a model of a virtual flow meter to verify the existing flow rate measurements and restore the missing values. The basis of the physical and mathematical approach is an algorithm that connects the parameters of the ESP system with the flow rate through the power consumption and hydraulic calculations of the gas-liquid mixture flow in the tubing. After preliminary calibration of pressure-flow characteristics of ESP for real mode of operation, daily measurements of fluid flow rate with a periodicity of 1 hour were calculated.

The accuracy and precision of well rates are paramount in reservoir management, well performance surveillance, flow assurance, and any third party processing arrangements. Rate allocation is traditionally based on well rate tests and downtime. This method is usually time-consuming and thus performed relatively infrequently. This could be inadequate for proactive asset management especially with wells that may produce in transient state. This paper discusses methodology to improve well rate allocation quality and save engineering time. In practice, many fields face some or all of the following challenges that are related to well rate allocation: 1) reservoir communication, 2) well interference, 3) changing skin factors or other near wellbore boundaries, 4) uncertainties around reservoir fluid properties, and 5) difficulty in obtaining well rate tests for a variety of reasons. For many intelligent wells, it is a common practice to install permanent downhole gauges which are playing critical roles in field management. This paper describes a framework on how to capitalize on the real time data from well pressure and temperature sensors and use the data in Integrated Asset Modeling (IAM) to allocate the well rate with enhanced accuracy, increased frequency, and reduced processing time. The paper uses Atlantis in Gulf of Mexico as an example to demonstrate this process. The real-time data supported model based allocation process becomes virtual flow meters for the intelligent wells. For Atlantis, field-wide allocation accuracy has been improved from previously +/− 10% error using the traditional allocation method based on well rate test and downtime method to current +/−3% error using the new method. This paper also shows model maintenance is a journey that needs strenuous attention especially after water injections commences and more production wells come online.

Virtual Flow Metering for Continuous Real-Time Production Monitoring of Unconventional Wells

The availability of a simple and robust flow allocation system is of primary importance for reservoir management since it provides oil, water, and gas production for each well. The low frequency of well separator tests and the difficulties in performing regular maintenance of multiphase flow meters have led to the development of Real Time Virtual Flow Meter (RTVFM) in Eni, a numerical solution to obtain real time flow rate estimation from pressure/temperature gauges measurements. This paper discusses the implementation and application of a novel RTVFM algorithm that increases the accuracy, stability, and robustness of the existing numerical tools even in case of extreme oil field environment with significant uncertainties. Current Virtual Meter algorithms are based on fluid dynamic simulators which calculate the pressure drops through wellbore, choke, and flowlines; the algorithm can be run in real time to find the optimal production rates that minimize the error between physical pressure readings and the calculated ones. In this work, a constraint is added to the system by including the temperature matching in the objective function, further improving the tool reliability. An accurate heat transfer characterization of the well has been implemented to predict the temperature changes along the wellbore during time, as well as the thermal effect due to pressure variations (Joule-Thompson effect). The effectiveness of the implemented algorithm has been proven by its application on a few offshore oil producers. In the chosen wells, equipped with dedicated MPFMs, the production measurements are not always reliable and RTVFM can be a valid support tool for back allocation. However, the flow rate estimation can be affected by significant uncertainties like production parameters variability (water cut and gas oil ratio) and fluid properties variation due to gas re-injection or artificial gas lift. In this scenario, the proposed enthalpy balance model allows to find a unique solution for the flow rate estimation, while the algorithm based only on pressure readings can converge to multiple solution rates. Increasing the accuracy of RTVFM tool is imperative to allow a reliable back allocation process, even in case of MPFM unavailability, poor sensors data quality and highly variable fluid properties. This paper investigated how an advanced thermo-fluid dynamic model can improve Virtual Meter algorithms, thus reducing the uncertainties in the numerical flow rate estimation.

Use of Trend Data from BEMS for the Ongoing Commissioning of HVAC Systems