Closed-loop reservoir management (CLRM) entails continuous data collection and the repeated application of history matching followed by production optimization. In traditional CLRM procedures, high-fidelity reservoir simulation is used for all production optimization and history matching computations. This approach can be very expensive in practice, particularly when robust optimization is applied (meaning optimization is performed over multiple realizations) and a derivative-free optimization procedure is used. Therefore, a proxy that can be applied to optimize well controls over multiple geomodels in an ensemble will be particularly useful. In recent work, we developed a long short-term memory recurrent neural network (LSTM RNN) proxy to perform well control optimization, under nonlinear output constraints (e.g., maximum well water cut, maximum field water production), for a deterministic geological model (Kim & Durlofsky, SPEJ, 2021). In this work, we extend this methodology to predict well-by-well production/injection rate time series for each geomodel in an ensemble, for a specified bottomhole pressure (BHP) schedule. This is achieved by incorporating a convolutional neural network (CNN) into the RNN-based proxy. In the CNN–RNN proxy, the CNN processes permeability realizations and provides initial long and short-term states for the RNN. The RNN accepts the initial states and specified BHP time-series and provides the well responses, allowing for the computation of expected NPV and constraint violations over the ensemble.

The CNN-RNN proxy is incorporated into a CLRM framework and applied to a 3D system characterized by Gaussian permeability realizations. The goal is to maximize NPV for a waterflood subject to maximum water injection rates and maximum water cuts. Production optimization is achieved using the CNN–RNN proxy with a particle swarm optimization algorithm, with a filter-based treatment for constraint handling. Permeability fields are represented using principal component analysis. History matching is accomplished using a randomized maximum likelihood method with gradient-based minimization. Computational results demonstrate that the CNN–RNN proxy accurately predicts well-by-well oil and water rates, for a wide range of input BHP profiles, for each realization in the ensemble. CLRM results show that NPV for a synthetic “true model” increases over the five closed-loop cycles considered. Substantial uncertainty reduction is also achieved as CLRM progresses.

One of the key challenges in mature fields is to maintain economic viability of production under varying reservoir and operating conditions. Artificial lift is considered an essential component on both sides of the equation as it “lifts both the revenue and cost”. Reactivating closed wells is another economic approach to improve profitability. The objectives of the work presented here are two-fold. We apply a novel reservoir modeling technique called Data Physics to optimize production under the existing pump conditions and get recommendations for pump rightsizing when required. Furthermore, we utilize the same model to reactivate the most profitable closed wells in the field.
Data Physics combines conventional reservoir physics with machine learning, data assimilation and advanced optimization techniques. In this paper, a Data Physics model is trained for a mature field and evolutionary algorithms were used to find optimal water injection plan (unconstrained) maximizing long-term and short-term production and NPV as well as minimizing water injection capacity. The optimization was performed on the injector side and does not necessarily address the fact that the resulting gross production may fall outside the optimum operational ranges of the down-hole pumps. To address this constraint, the optimization tool was then modified to include bottom-hole pressure and down-hole pump operational parameters and a new plan (constrained) was generated. The new constraints are nonlinear functions of the optimization parameters and there are as many of these constraints as the number of wells with down-hole pumps which makes the problem challenging for the optimization algorithm.
The unconstrained redistribution plan (maintaining the same water capacity) led to an increase of 3.9% in production while the constrained plan resulted in 2.2% in production. This comparison allowed the operator to estimate the potential incremental oil and the added value created by replacing existing pumps by more appropriated pumps and to perform economic analysis of the NPV of such replacements. Additionally, a simplistic one-at-a-time reactivation optimization was performed to identify the top 10 wells in terms of production. Reactivating these wells would lead to an expected 11.7% further gain in production.

Monitoring of the CO2 plume location is an essential component of any carbon storage project. The optimal placement of monitoring wells is challenging because this must be accomplished before CO2 injection is started, when geological uncertainty is high. In addition, this optimization is closely linked with the history matching procedure, as the measurements from the monitoring wells represent the key observations used for data assimilation. In this work, we present and apply a framework that integrates the monitoring-well optimization and history matching problems. Our approach is ensemble-based and is expressed within a Bayesian setting. The monitoring-well optimization entails finding the locations of monitoring wells such that, with the data acquired at those locations, the expected variance of a quantity of interest is minimized (related approaches were used previously by He at al., 2018; Sun & Durlofsky, 2019, in data-space inversion settings). This quantity of interest, which must be aligned with the goals of the history matching, is here taken to be the volume of CO2 beyond a specified distance from the injector. Through use of prior simulation data and a genetic algorithm-based minimization, we thus find the locations of monitoring wells such that, when the data from these wells are used in history matching, we maximize uncertainty reduction.

The overall framework is applied to idealized multi-Gaussian geomodels based on an actual storage project under development in the US. A large number of prior simulation runs are required for monitoring-well optimization, and these are conducted using an existing deep-neural-network surrogate model, developed by Wen et al. (2021). Following this optimization, history matching is performed using an ensemble smoother with multiple data assimilation. Several different (synthetic) “true” models, which provide the observed data, are considered. The requisite history matching simulations are again accomplished using the surrogate flow model. We generate history matching results for optimal monitoring-well locations and for heuristically or randomly placed monitoring wells. In all cases posterior uncertainty, evaluated in terms of the cumulative distribution function for plume extent over all history-matched models, is found to be minimized through use of optimized monitoring wells. This confirms the consistency and applicability of the integrated workflow. We note finally that the overall framework is general, and it is compatible with different optimization procedures, history matching methods, and flow simulators.