Closed-Loop Feedback Control for Production Optimization of Intelligent Wells Under Uncertainty F.A. Dilib, SPE, and M.D. Jackson, SPE, Imperial College London Summary A key component of intelligent-well systems is the decision framework used to identify control actions for production optimi- zation. Most published algorithms use model-based control, yet model-based techniques are effective only if the model or ensem- ble of models used in the optimization captures all possible reser- voir behaviors at the individual-well and -completion level. This is rarely the case. Moreover, reservoir models are rarely predic- tive at the spatial and temporal scales required to identify control actions. We show that simple, closed-loop feedback control, trig- gered by monitoring at the surface or downhole, can increase the net present value (NPV) and mitigate reservoir uncertainty. We do not neglect reservoir-model predictions entirely; rather, we use a model-based approach to optimize adjustable parameters in feedback-control strategies. We compare open-loop control, by use of fixed control devices (FCDs), with closed-loop feedback control, by use of “on/off” inflow-control valves (ICVs), operated in response to monitoring at the wellhead, and “variable” ICVs operated in response to mon- itoring downhole. We also use a gradient-based optimization algorithm to find the dynamic optimal inflow-control behavior. This strategy assumes perfect reservoir knowledge and is imple- mented only for benchmarking of the feedback-control strategies. Our result suggests that closed-loop control, on the basis of direct feedback between reservoir monitoring and inflow-valve settings, can yield close-to-optimal gains in NPV compared with uncontrolled production, even if the reservoir behavior lies out- side the range predicted by reservoir models. Moreover, similar gains are observed by use of surface monitoring and simple on/off ICVs, and downhole monitoring and variable ICVs. In contrast, open-loop control yields significantly lower NPV gains, and is also a riskier strategy because unpredicted reservoir behavior can yield suboptimal sizing of the FCDs and negative returns. Introduction Intelligent (or smart/advanced) wells are equipped with wellhead and/or downhole sensors to monitor well and reservoir conditions, and with FCDs or variable ICVs to control the inflow of fluids from the reservoir (Robison 1997). Both control devices impose an addi- tional pressure drop between the sandface and production tubing, but FCDs are fixed and must be sized before installation, whereas ICVs can be varied by the operator during production. The deploy- ment of intelligent wells has been shown to improve oil recovery in a range of reservoir, well, and production scenarios (e.g., Algeroy et al. 1999; Glandt 2005); nevertheless, the development of robust control techniques to identify the optimal FCD or ICV settings is still an area of active research, particularly when there is uncertainty associated with the predictions of the reservoir model. Numerous previous studies have demonstrated that open- and closed-loop con- trol strategies can add significant value (e.g., Brouwer et al. 2001; Yeten and Jalali 2001; Brouwer and Jansen 2004; Yeten et al. 2004; Aitokhuehi and Durlofsky 2005; Elmsallati and Davies 2005; Elm- sallati et al. 2005; Sarma et al. 2005a; Saputelli et al. 2005; Ebadi and Davies 2006; Naus et al. 2006; Doublet et al. 2009; Jansen et al. 2009; Datta-Gupta et al. 2010). Open-loop control strategies are required with FCDs because it is not possible to modify the device settings after installation without a major and expensive well work- over; closed-loop control strategies are possible in wells equipped with ICVs and surface and/or downhole monitoring. Most previous studies have relied on the predictions of reservoir and well models to identify the optimal FCD/ICV settings. Yet reservoir models are rarely predictive over the spatial and temporal resolution required for inflow-control decisions. Key aspects of flow may be controlled by fine-scale geological features that are below model resolution or cannot be located spatially with confidence (e.g., thin, high-perme- ability intervals or mudstone barriers); production may also be dominated by near-well effects such as cusping or coning, which are often poorly captured in simulation models. Furthermore, the predictions of reservoir models are typically associated with signifi- cant uncertainty; even history-matched models can lack predictive value (e.g., Tavassoli et al. 2004; Carter et al. 2006). The ability of simulation models to predict gross or average reservoir behavior successfully has been demonstrated in numerous studies, but so has their failure to predict the detailed aspects of individual-well and - completion behavior, which is necessary for day-to-day inflow-con- trol decisions (e.g., Gringarten et al. 2003; Daltaban et al. 2008). The aim of this paper is to investigate whether simple closed- loop control strategies, on the basis of direct feedback between ICV settings and surface or downhole measurements, can enhance well NPV and mitigate reservoir uncertainty. An earlier study by Addiego-Guevara et al. (2008) shared a similar aim; however, our approach differs from previous studies in a number of ways. First, we develop closed-loop, direct-feedback control strategies that are designed to work across a range of reservoir and production scenarios. Previous studies have investigated direct-feedback con- trol, but the control actions were identified on an ad hoc basis (e.g., Brouwer et al. 2001; van der Poel and Jansen 2004; Elmsal- lati et al. 2005; Grebenkin and Davies 2010); here, we develop a general feedback relationship that contains a number of adjustable parameters, and optimize the value of these parameters using res- ervoir-model predictions. Thus, we do not omit model predictions entirely; however, rather than basing control actions directly on model predictions, we use model predictions to optimize a direct- feedback control relationship between measured data and inflow- control settings. Second, we test the robustness of our direct-feed- back control algorithms specifically against unexpected reservoir behavior. Previous studies applied model-based inflow control by use of a range of reservoir model realizations, but assumed that the realizations captured all possible production behaviors at the individual-well and -completion level. In practice, this is usually not the case, for the reasons discussed previously. Here, we simu- late production using models that lie outside of the range used to optimize the adjustable parameters in our direct-feedback control relationships, to mimic the application of direct-feedback control when the reservoir does not behave as predicted. Third, we bench- mark the performance of our feedback-control methods against a model-based approach that assumes perfect reservoir knowledge. We argue that such an approach is rarely possible in practice, but here it allows us to determine whether direct-feedback control yields close-to-optimal solutions. Previous studies investigated Copyright V C 2013 Society of Petroleum Engineers This paper (SPE 150096) was accepted for presentation at the SPE Intelligent Energy International, Jaarbeurs, Utrecht, the Netherlands, 27–29 March 2012, and revised for publication. Original manuscript received for review 26 July 2013. Paper peer approved 30 July 2013. November 2013 SPE Production & Operations 345