Automatic prediction of reservoir inflow using data-driven physical modelling

Norway is the world leading hub in multiphase flow research and innovation, and, in 2012 multiphase flow technology was elected one of two most important technologies for the Norwegian society since 1980. This research has resulted in renowned multiphase flow simulators, such as OLGA, LedaFlow and now the new simulator from Turbulent Flux. Today these multiphase flow simulators play an important role in both the design and daily operations of most oil & gas fields worldwide. While these simulators are routinely used to give valuable insights into production operation, they are rarely used to offer real-time advice to the operators. Instead, these tools are used reactively by a small group of in-house experts to recommend mitigations. Turbulent Flux is dedicated to making these simulations more readily available for giving real-time advice to the people in the control room. To achieve this, a crucial step is to both reduce the complexity of use and effort to maintain such systems. One of the main issues in using and maintaining real-time flow-simulation systems is to determine the reservoir inflow into a well. Especially when the inflow into the well changes over time. The reservoir inflow is used as a boundary condition in the flow-simulation and is critical for obtaining good results. Usually, this is modelled using semi-empirical mathematical models which are calibrated to a given flow situation. However, as these relationships do not account for the transients involved in reservoir inflow, they often require continuous tuning and updating. The current digitalization trend in the oil & gas industry makes much more data available, data which can be utilized to automatically determine the unknown parameters in the inflow models in real time. In this project, we have focused on using a combination of flow simulation models and data analysis to automatically determine which models best describe reservoir behavior and the unknown parameters in these models. This reduces the complexity of real-time multiphase simulations and makes these systems maintenance free. A central task in the project has been to group and classify existing semi-empirical mathematic models for reservoir inflow. The execution of this task has been based on a combination of measurements and synthetic data generated by simulations. In addition, the project has conducted a thorough review of published data-driven methods for reservoir inflow without discovering any works of note. As an alternative to the predominantly steady-state based IPR models, the project has developed new transient models for pressure and flow rates. This development was required when the data revealed a lack of IPR-type correlation. The models were developed in order to capture both transient and steady-state conditions, based on conservation of mass and Darcy-type relations of pressure gradient and flow rate. Data analysis has been used to determine parameters in published semi-empirical IPR models, and in order to rank them according to performance. This has been performed by using historical data to develop models for prediction of future flow rates. This has resulted in two applications, one based on direct regression and one based on Bayesian approach. The purpose of both of these applications are to: 1. Establish model parameters from historical data 2. Estimate and rank the quality, model fit and robustness 3. Estimate the temporal development of model parameters The project results show that published reservoir inflow models have clear limitations, especially when it comes to their ability to predict into the future. The project has concluded that the free parameters in these models, combined with the uncertainty in the measured data, makes it impossible to determine all model parameters with a high degree of confidence, when the reservoir response changes significantly. The project has resulted in improved and continuously updated reservoir inflow models. However, the validity of these models is limited to near-future predictions only. The developed methods and applications can be used to validate and improve IPR-models in commercial use, thus resulting in improved predictions. This will result in more reliable results of virtual flow metering solutions, enabling the use of these to enhance the efficiency and safety in existing oil and gas production systems.

Få publiserte studier har brukt dataanalyse for å bestemme frie parametere i innstrømningsmodeller (IPR). Disse er sentrale for operatørers produksjonsplanlegging. Forbedringer i IPR bidrar indirekte til mere effektiv og sikker produksjon og har derfor stor verdi for operatør. Gjennom prosjektet har vi vist at publiserte IPR har klare begrensninger, resultatene kan brukes til å øke bevisstheten om gyldighetsområdet. Metodene som er utviklet ved å kombinere dataanalyse av brønndata og IPR brukes for å validere og forbedre modellene. Integrert i virtuelle strømningsmålere gir dette mer pålitelige resultater. Turbulent Flux vil jobbe videre med å tilpasse metodene, med mål om å bli fullt integrert i kommersielle løsninger, samt presentere i forum relevante for industrien. Prosjektet har involvert parter fra ulike domener, med bakgrunn i fysikk, matematikk, programvareutvikling og industri. Dette har bidratt til å etablere et fagmiljø i Norge med ny kompetanse som det kan bygges videre på.

Norway has been the world leading hub for multiphase research and innovation for 40 years. This has resulted in renowned transient multiphase flow simulators, such as OLGA, LedaFlow, FlowManager and now the simulator from Turbulent Flux. These simulators are ubiquitous and play an important role in both the design and daily operations of oil and gas fields. A key driver in the current digitalization trend in the industry is the desire to leverage vast amounts of data to support and guide operational decisions. This includes using sensor data to guide real-time multiphase flow simulations in understanding physical processes in production systems. One such application is Virtual Flow Meters (VFM). A significant challenge in using this type of technology in e.g. VFM, is to accurately predict the flow from the reservoir into the well. In fact, modelling the proper response from the reservoir is a key component to predict flow-behaviour ahead in time. Although existing solutions provide approaches to specify the reservoir inflow, they all rely on engineering experience and regular intervention and maintenance. It is industry practice to use Inflow Performance Relationships (IPR’s) to model reservoir inflow. A multitude of IPR models exists, but all are valid only within a limited parameter range. Moreover, selecting an IPR model is a manual task, left to the judgement of the engineer. Furthermore, all models contain empirical constants, which are manually updated as the conditions in the reservoir changes. With the increased availability of sensor data, it is now possible to take a new and data-driven approach to predict reservoir inflow. The project will use large amounts of data, both system- and sensor-data, to guide the determination of inflow. This approach will increase the accuracy of real-time multiphase flow simulations, reduce installation cost and remove costly maintenance of the system completely.