Simulation ensemble-based well placement optimization

Abstract

A computer-implemented method, a system, a computer program, and a storage medium for optimizing well placement using artificial intelligence / machine learning (AI / ML) models (i.e., metamodels). The computer-implemented method comprises the steps of: performing a first process using at least one field-level metamodel to select field control and well configuration parameters, the field-level metamodel being adapted to mimic field behavior (110); performing a second process using at least one well-level metamodel to predict a well key performance indicator (KPI), the well-level metamodel taking into account downhole completion data and preferably well-level production and / or injection data, and field control and well configuration parameters selected from the first process (120); generating an opportunity map based on results from the second process (130); enhancing the opportunity map by incorporating physical constraints on well drilling locations (140); and determining well placement based on the enhanced opportunity map (150).

Description

Simulation Ensemble-based Well Placement OptimizationField[oooi] The present invention generally relates to well placement optimization in field development planning and, more specifically, to a computer-implemented method, a system, a computer program, and a storage medium for optimizing well placement using artificial intelligence / machine learning (AI / ML) models. The AI / ML models are also referred to as metamodels in the present disclosure.Background

[0002] In the oil and gas industry, well placement planning is used to select placements for drilling wells into a subsurface reservoir to reach specific locations in the reservoir that are believed to contain recoverable hydrocarbons. Well placement planning may include determining the location of one or more wells, as well as additional information such as well trajectories, well completions, drilling schedules, etc.

[0003] Generally, a reservoir simulator is used in connection with well placement planning. A reservoir simulator is a computational tool used to model and predict the behavior of oil and gas reservoirs over time. The process involves creating a mathematical model of the reservoir based on geological and engineering data. Once the mathematical model is created, it is used to simulate the behavior of the reservoir under different conditions.

[0004] Conventionally, in well placement planning, a user selects target and well locations, performs a reservoir simulation forecast, and then calculates a value based on the forecast oil or gas recovered and the cost of the wells. The user may repeat the process while modifying the number and location of the wells and completions. The modifications may further include different water flooding strategies, well spacing, well types, platform locations, etc.

[0005] Well placement planning is known to be an overly complex and time-consuming task. A more effective and computationally efficient approach to well placement planning has been the topic of many publications.[ooo6]US 10,060,226 B2 (or US 2015 / 0233214 Al) discloses a well design system that utilizes geological characteristics and fracture growth behavior along of a vertical stratigraphic column of the formation in order to optimize well placement and fracture stimulation designs for the entire formation.

[0007] US 8,155,942 B2 (or US 2009 / 0216505 Al) discloses methods, systems, and software for optimizing well placement in a reservoir field. The locations of the wells are represented by continuous well location variables associated with a continuous spatial domain. A gradient of the objective function is calculated responsive to the continuous well location variables. The locations of the wells are then adjusted responsive to the calculated gradient of the objective function. Iterative calculation of the gradient and adjustment of the wells continue until the well locations are optimized.[ooo8]EP 2 948 618 Bi (or WO 2014 / 117030 Al) discloses a method, apparatus, and program product that utilize a constrained optimization framework to generate a well placement plan based on a reservoir model.

[0009] US 2020 / 0080406 Ai discloses systems, methods and apparatuses for a Black Hole Particle Swarm Optimization for Optimal Well Placement in Field Development Planning.

[0010] US 7,478,024 B2 (or US 2005 / 0149307 Al) discloses a method of managing a fluid or gas reservoir. The method assimilates diverse data having different acquisition time scales and spatial scales of coverage for iteratively producing a reservoir development plan that is used for optimizing an overall performance of a reservoir.

[0011] US 8,005,658 B2 (or US 2008 / 0300793 Al) discloses a hybrid evolutionary algorithm (“HEA”) technique for automatically calculating well and drainage locations in a field. The technique includes planning a set of wells on a static reservoir model using an automated well planner tool that designs realistic wells that satisfy drilling and construction constraints. A subset of these locations is then selected based on dynamic flow simulation using a cost function that maximizes recovery or economic benefit.

[0012] US 8,793,111 B2 (or US 2010 / 0185427 Al) discloses a system for automatically optimizing a Field Development Plan (FDP) for an oil or gas field uses a fastanalytic reservoir simulator to dynamically model oil or gas production from the entire reservoir over time.

[0013] WO 2020 / 256738 Ai discloses a method for determining an action associated with field development planning or asset management of an oilfield based on a policy or value function and a state, simulating a reward of the action using information identifying the action and the state as inputs, generating an experience record based on executing the simulation, wherein the experience record identifies the simulated reward associated with the action and a current state or resulting state, storing the experience record in an experience buffer, wherein the experience buffer stores a plurality of previously generated experience records, updating the policy or value function based on the information stored in the experience buffer, selecting an action related to the field development planning based on the updated policy, and displaying information representing the action.

[0014] Document “A Field Development Strategy for the Joint Optimization of Flow Allocation, Well Placements and Well Trajectories” (Energy Exploration & Exploitation; Volume 39, Issue 1; Pages: 502 - 527) discusses optimal flow settings and well locations in order to set a strategy for the development of a virtual offshore oil field and presents a well trajectory optimization methodology.

[0015] Document “Practical Closed-Loop Reservoir Management Using Deep Reinforcement Learning” (Nasir, Yusuf, and Louis J. Durlofsky; SPE J. 28 (2023): 1135-1148. doi: https: / / doi.org / 1o.2118 / 212237-PA) discusses a closed-loop reservoir management (CLRM) framework. This framework uses deep reinforcement learning (DRL) to train control policies that directly map observed well data to optimal well settings.

[0016] In summary, most of the existing optimization techniques rely on reservoir simulations in the prediction and optimization process. In addition, the optimization is done iteratively through interaction with the reservoir simulator. The existing optimization techniques are still very time consuming and expensive. Therefore, it is a goal for the present invention to provide more efficient techniques for well placement optimization.Summary

[0017] The invention is defined by the independent claims. Preferred embodiments are defined in the dependent claims.

[0018] The various embodiments detailed herein relate to a computer-implemented method, a system, a computer program, and a storage medium for well placement optimization. The embodiments and aspects of the present invention provide other benefits that will become clear to those skilled in the art from the foregoing description.

[0019] An example computer-implemented method may comprise the steps of performing a first process using at least one field-level metamodel to select field control and well configuration parameters, the field-level metamodel being adapted to mimic field behavior; performing a second process using at least one well-level metamodel to predict a well key performance indicator (KPI), the well-level metamodel taking into account downhole completion data and preferably well-level production and / or injection data, and field control and well configuration parameters selected from the first process; generating an opportunity map based on results from the second process; enhancing the opportunity map by incorporating physical constraints on well drilling locations; and determining well placement based on the enhanced opportunity map.

[0020] An example computer program may comprise computer readable instructions which, when implemented on a computer, cause the computer to carry out the following steps: performing a first process using at least one field-level metamodel to select field control and well configuration parameters, the field-level metamodel being adapted to mimic field behavior; performing a second process using at least one well-level metamodel to predict a well key performance indicator (KPI), the well-level metamodel taking into account downhole completion data and preferably well-level production and / or injection data, and field control and well configuration parameters selected from the first process; generating an opportunity map based on results from the second process; enhancing the opportunity map by incorporating physical constraints on well drilling locations; and determining well placement based on the enhanced opportunity map.

[0021] An example system may comprise a processor; a memory coupled with the processor; and instructions stored in the memory and executable by the processor to cause the system to perform the following steps: performing a first process using at least one field-level metamodel to select field control and well configuration parameters, the field-level metamodel being adapted to mimic field behavior; performing a second process using at least one well-level metamodel to predict awell key performance indicator (KPI), the well-level metamodel taking into account downhole completion data and preferably well -level production and / or injection data, and field control and well configuration parameters selected from the first process; generating an opportunity map based on results from the second process; enhancing the opportunity map by incorporating physical constraints on well drilling locations; and determining well placement based on the enhanced opportunity map.

[0022] Another example system may comprise means for performing a first process using at least one field-level metamodel to select field control and well configuration parameters, the field-level metamodel being adapted to mimic field behavior; means for performing a second process using at least one well-level metamodel to predict a well key performance indicator (KPI), the well-level metamodel taking into account downhole completion data and preferably well-level production and / or injection data, and field control and well configuration parameters selected from the first process; means for generating an opportunity map based on results from the second process; means for enhancing the opportunity map by incorporating physical constraints on well drilling locations; and means for determining well placement based on the enhanced opportunity map.Brief description of the drawings

[0023] The accompanying drawings are included to provide a further understanding of the invention. The drawings illustrate embodiments of the invention and, together with the description, serve to explain the principles of the invention.

[0024] In the drawings:

[0025] FIG. 1 illustrates an example method for determining well placement according to aspects of the disclosure.

[0026] FIG. 2 illustrates an example method for training a field-level metamodel according to aspects of the disclosure.

[0027] FIG. 3 illustrates an example method for selecting optimal field control and well configuration parameters according to aspects of the disclosure.

[0028] FIG. 4 shows an example of opportunity map generated based on results from a well-level metamodel according to aspects of the disclosure.

[0029] FIG. 5 shows an example of enhanced opportunity map according to aspects of the disclosure.

[0030] FIG. 6 shows an example of opportunity map added with optimal well placement locations according to aspects of the disclosure.Detailed description

[0031] The following detailed description describes techniques for well placement optimization using AI / ML models. The present disclosure is not intended to be limited to the described or illustrated examples, but to be accorded the widest scope consistent with the described principles and features.

[0032] To facilitate the optimization for well placement in field development planning, the present disclosure provides techniques which utilize AI / ML models. The AI / ML models are also referred to as metamodels in the present disclosure. The techniques disclosed herein include a hierarchical data handling workflow that operates at the field, group, well, and / or completion levels. Various metamodels are trained to estimate field development outcome at the above-mentioned levels.

[0033] The techniques disclosed herein can be used to plan optimized management of subsurface assets by performing economic and recovery-based optimization of field activities. This optimization is achieved through well placement and scheduling to meet business objectives. More specifically, those techniques provide for the optimization of well placement and design to maximize the recovery of hydrocarbons, while considering operational constraints and minimizing environmental impacts.

[0034] Overall, the use of metamodels allows for rapid evaluation of field control scenarios and the optimization of well placement and design to achieve business objectives. The techniques disclosed herein can be used in a wide range of oil and gas reservoirs and can be adapted to changing conditions in the reservoir over time.

[0035] The techniques disclosed herein provides a loosely coupled system between the field level and the well level. A metamodel is first built at the field level to optimize field control parameters such as facility capacities (e.g., how much oil, water / liquid and / or gas volumes are to be processed in a day) and well operating control conditions (e.g., maximum production / injection rates, maximum / minimum pressure to operate wells, etc.). The facility capacities act as constraints to dailyproduction affecting cash flow. The capacities are also key to calculating Capital Expenditure (CAPEX). Well operating control conditions may also include heuristic economic measures to close production wells based on water-cut and gas-oil-ratio. The field level metamodel is kept unaware of details of well locations but has features related to well spacing. Then, downhole completion perfo ration / completion data and well-level production and injection data are merged. The completion data include the geolocation and permeability-length characteristics of the well connections. A key performance indicator (KPI) is calculated based on production and injection data. The field-level control parameters selected by the field level metamodel are then added to create a welllevel metamodel. The KPI can be e.g., integrated cumulative production / inj ection volumes (e.g., integrated Well Oil Production Total (iWOPT) / integrated Well Water Injection Total (iWWIT) / integrated Well Gas Injection Total (iWGIT), which is the area under the curve for oil production / water / gas injection totals). Such KPI prioritizes wells on production / injection volumes and how fast it is obtained. The use of such kind of KPI favors wells economically in terms of “time value of money”.

[0036] Unlike conventional well optimization methods, the techniques disclosed herein rely mainly on simulation results rather than static input parameters (e.g., geological properties such as porosity, permeability, or fluid saturations map, etc.). The metamodels are trained on numeric simulation output data, which can be considered as dynamic results. The dynamic results may include, for example, well production and injection data reflecting combinations of different field control and different well location and spacing parameters. The dynamic results can reflect the quality of rock properties where the well is located. The dynamic results can also be influenced by well interference and dynamic field pressure changes during the life of the field. Therefore, the metamodels implicitly consider both static and dynamic factors determining the success of the production and injection wells.

[0037] The techniques disclosed herein will be better understood with reference to the implementation examples discussed below.

[0038] FIG. 1 illustrates an example method for determining well placement according to aspects of the disclosure. The well placement includes, for example, well locations and / or well scheduling, etc. The method can be implemented by any suitable means. For example, the method can be implemented by a systemcomprising at least one processor and at least one memory. The at least one memory comprises instruction, which, when executed by the at least one processor, cause the at least one processor to carry out one or more steps of the method disclosed herein. The system can be partly or completely set up in the cloud.

[0039] As shown in FIG. 1, at block 110, the system or any suitable means performs a first process using at least one field-level metamodel to select field control and well configuration parameters. The field-level metamodel is adapted to mimic field behavior. The field-level metamodel can be also understood as an asset-level AI / ML model or metamodel.

[0040] In some implementations, the field-level metamodel may be modeled to determine a net present value (NPV), ultimate total recovery (UTC), or any other economic key performance indicator (KPI). For example, a set of field control and well configuration parameters can be selected based on the criterion of maximizing the NPV, maximizing the hydrocarbon recovery, and / or minimizing the cost.

[0041] In some implementations, the field-level metamodel may be constructed based on Powell method and / or Particle Swarm Optimization method.

[0042] As further shown in FIG. 1, at block 120, the system or any suitable means performs a second process using at least one well-level metamodel to determine a well KPI such as integrated cumulative production / inj ection volumes (e.g., integrated well oil production total (iWOPT), integrated well water injection total (iWWIT), integrated well gas injection total (iWGIT), etc.).

[0043] As further shown in FIG. 1, at block 130, the system or any suitable means generates an opportunity map based on results from the second process. An example opportunity map generated at block 130 is shown in FIG. 4.

[0044] In some implementations, the opportunity map may be generated to represent a whole reservoir. In some other implementations, multiple opportunity maps may be generated, each representing a respective zone in the reservoir.

[0045] As further shown in FIG. 1, at block 140, the system or any suitable means enhances the opportunity map by incorporating physical constraints on well drilling locations. An example of enhanced opportunity map, generated at block 140, is shown in FIG. 5.

[0046] In some implementations, the physical constraints on well drilling locations may include geological features. The geological features may represent areas in the reservoir that cannot and / or should not be drilled. In some implementations, the geological features comprise faults and / or existing well locations, etc.

[0047] As further shown in FIG. 1, at block 150, the system or any suitable means determines well placement based on the enhanced opportunity map. An example of opportunity map added with optimal well locations, generated at block 150, is shown in FIG. 6.

[0048] In a further aspect, the opportunity map may be generated for multiple timesteps. In one example implementation, each timestep may have a duration of one year.

[0049] In a further aspect, the existing well locations may be updated when one or more new wells have been drilled. In some implementations, the update takes care of previously selected areas for drilling, such that they will become unavailable in subsequent optimization iteration, wherein one iteration may correspond to one timestep.

[0050] In a further aspect, well trajectories and schedule for reservoir simulation may be exported for validation. For example, the well schedules may be generated through a well-count model, which is adapted to determine a number of drilled wells and a number of abandoned wells at each iteration. The well trajectories may be a combination of individual well locations as an output of the well optimization workflow (e.g., shown in Figure 1, block 150). In one example implementation, the well trajectories may further include information about well configurations of the individual wells. In one example implementation, validation can be done by running the aforementioned exported data in a physics-based simulator and comparing results from the physics-based simulator with results from the metamodels.

[0051] In a further aspect, the field-level metamodel and / or well-level metamodel may be trained based on simulation results from a plurality of simulation cases with various field control and well configuration scenarios. The simulations can be also referred to as ground-truth simulations. An example implementation of training the field-level metamodel is shown in FIG. 2.

[0052] In one example implementation, each field control and well configuration scenario may comprise a respective set of field control parameters and well configuration parameters. For example, the field control parameters may include parameters relating to production facilities (e.g., oil production capacity, water / liquid production capacity, gas production and injection capacity, gas lift capacity, etc.) and parameters relating to well management (e.g., oil / liquid / gas production rate, water / gas injection rate, production / injection operating Bottom Hole Pressure (BHP), maximum water-cut to shut-in the well, maximum gas-oil ratio (GOR) to shut-in the well, etc.); and the well configuration parameters may include parameters relating to well spacing, well pattern, a number of production and / or injection wells, etc.

[0053] In a further aspect, an optimal set of field control and well configuration parameters can be selected based on a forward optimization process, such as shown in FIG. 3.

[0054] In some implementations, the metamodels may be constructed based on one or more machine learning approaches, e.g., neural network, XGBOOST, RandomForest, etc.

[0055] The techniques explained with reference to Fig. 1 offer significant advantages over the prior art, such as not requiring access to a reservoir simulator during the optimization process and utilizing simulation results as the main input instead of 3D grid properties or well logs. Those techniques can improve field development planning activities and optimize subsurface asset performance, resulting in increased production and economic benefits.

[0056] Since the optimization process can be performed without access to the reservoir simulator, the simulations and optimizations can be conducted in different systems. For example, the simulations can be run on a complex system with sufficient computation power while the optimizations can be run on a budget system with less computation power, since the optimization step according to the present invention is less computation demanding compared to the simulation step.

[0057] Furthermore, the metamodels generated based on the techniques disclosed herein can learn static properties’ influence and dynamic effects, such as well interference and pressure changes, directly from simulation results. This means theoptimization process does not need to decide on static and dynamic properties cutoffs, which can be arbitrary in conventional well-location optimization methods.

[0058] In addition, the workflow disclosed herein (such as shown in Figure 1) allows for integration of geological constraints into the opportunity map to make better informed optimization of well locations.

[0059] Besides, the techniques disclosed herein can be implemented not only to optimize well locations but also to optimize field level control parameters, leading to more comprehensive optimization of subsurface asset performance, considering both economic and recovery-based metrics.

[0060] FIG. 2 illustrates an example method for training a field-level metamodel according to aspects of the disclosure. As shown in Fig. 2, the field-level metamodel comprises two modeling components shown as 25 and 26. Although the field-level metamodel in this example comprises two components, it is possible to build the field-level metamodel with a different number of components or as one single component.

[0061] As shown in FIG. 2, the field-level metamodel has four types of input data, shown as 21, 22, 23, 24, and two types of output data, shown as 27, 28.

[0062] Input data 21 may include field control and well configuration parameters used for a set of simulation scenarios. For example, the field control parameters may include parameters relating to production facilities (e.g., oil production capacity, water / liquid production capacity, gas production and injection capacity, gas lift capacity, etc.) and / or parameters relating to well management (e.g., oil / liquid / gas production rate, water / gas injection rate, production / injection operating Bottom Hole Pressure (BHP), maximum water-cut to shut-in the well, maximum gas-oil ratio (GOR) to shut-in the well, etc.), etc. The well configuration parameters may include parameters relating to well spacing and / or well pattern.

[0063] Input data 22 may include parameters relating to physical properties of the wells used for the set of simulation scenarios. The physical properties maybe associated with information such as spatial location of well top, depth of the well top, perforation length, Toe-to-Heel distance, Northing extension, Easting extension etc.

[0064] Input data 23 may include information relating to static properties, such as porosity, permeability, etc., of the reservoir. In some implementations, input data 23 may be constructed in terms of 2D or 3D cubes, each cube describing a distribution of static properties of a respective region in the reservoir.

[0065] Input data 24 may include porosity, permeability information corresponding to each well. In some implementations, input data 24 may be obtained based on input data 22 and input data 23. For example, the input data 23 may be a 2D or 3D grid representing static properties of the reservoir in space. The intersection of the 3D grid with the well trajectories can be extracted based on well trajectories information in input data 22. The information extracted may form the porosity and permeability at the well level.

[0066] Output data 27 may represent a field-level output. In one example implementation, the output data 27 may be constructed as a field output curve, for example, cumulative oil production (OPT), cumulative water injection (WIT), etc.

[0067] Output data 28 may include a number of production / injection wells, which is the same as activation and abandonment time for production / injection wells.

[0068] The component 25 is trained to take input data 21 and input data 22 and to output corresponding output data 27, as shown in FIG. 2. The data used for training may be obtained based on simulation results from a physics-based simulator. In one example implementation, the component 25 may be a metamodel constructed based on one or more machine learning approaches, e.g., neural network, Extreme Gradient Boosting (XGBOOST), RandomForest, etc. For example, for neural network, component 25 may be trained by adjusting its internal parameters (weights and biases) during training using labelled data, minimizing the difference between its predicted outputs and the actual targets through a process called backpropagation and gradient descent optimization. XGBOOST is a boosting algorithm that sequentially trains decision trees to correct errors made by previous trees, optimizing a loss function. It combines these weak learners into a strong predictive model. RandomForest is an ensemble learning method that trains multiple decision tree models on random subsets of data and combines their predictions by averaging (for regression) or voting (for classification).

[0069] The component 26 is trained to take input data 21 and 24 and to output corresponding output data 28, as shown in FIG. 2. The data used for training maybe obtained based on simulation results from a physics-based simulator. In the physics-based simulators, wells are activated and abandoned (terminated) based on some local conditions (for example local saturation and pressure of oil) and some filed level condition (when total production capacity is reached some wells are shut down, when the total production is below the capacity new wells are activated). In other words, component 26 may be called a well-count model.

[0070] In a further aspect, the method of FIG. 2 may comprise a step of generating a large set of simulation cases with various field control and well configuration scenarios for training the field-level metamodel.

[0071] FIG. 3 illustrates an example method for selecting optimal field control and well configuration parameters according to aspects of the disclosure.

[0072] As shown in FIG. 3, a metamodel 32 (such as a field-level metamodel trained by a method described with reference to FIG. 2) is used in the optimization process.

[0073] The metamodel 32 takes input data 31 and produces data 33. The input data 31 may include a set of field control parameters, well-related parameters, and information on static properties, similar as input data 21, 22, and 23 as described in connection with FIG. 2. Among these, the field control and well parameters are optimized during the optimization process and are also referred to as optimization parameters. The data 33 may include field-level production curves and well timings (activation and abandonment) similar as output data 27 and 28 described in detail in connection with FIG. 2. For example, initially, a set of parameters as input data 31 may be chosen by a reservoir engineer or other experts of the field, based on a problem formulation.

[0074] The optimization process shown in FIG. 3 also uses an economic model 34. The economic model 34 takes data 33 as input data and generates an economic-related value 35. Value 35 represents an objective value that should be optimizes during the optimization process.

[0075] In some implementations, the economic model 34 may be defined through a set of mathematical functions that take field production and injection curves (such as the output data 27) and information on well timings (such as the output data 28) and return outputs such as net present value (NPV), discounted net present value, or ultimate total recovery (UTC), etc. In some other implementations, the economicmodel 34 may be constructed to consider additional information, for example, information on the type of each well (e.g., producer well, injector well, water injector, gas injector, or water-alternating-gas (WAG), etc.). For example, a NPV in the oil production context can be determined based on the following equation:

[0076] NPV = S [(R - C) / (1 + r)At]

[0077] Where: NPV = Net Present Value

[0078] S = Summation (from t=o to the project's end)

[0079] R = Net cash inflow (revenue) at time t

[0080] C = Net cash outflow (cost) at time t

[0081] r = Discount rate (the rate of return required for the investment)

[0082] t = Time period (usually measured in years).

[0083] In the above equation, the cash flows for each year (t) of the project's life are summed up. The cash flows include revenues (R) and costs (C) for each year, and they are discounted back to their present value using the discount rate (r).

[0084] In a more detailed example, for evaluating an oil production project that has the following cash flows:

[0085] Year o: Initial investment of $1,000,000[oo86]Year 1: Revenue of $500,000 and operating cost of $100,000

[0087] Year 2: Revenue of $600,000 and operating cost of $120,000[oo88]Year 3: Revenue of $700,000 and operating cost of $140,000

[0089] Assuming a discount rate of 10%, the NPV can be determined as follows:

[0090] NPV = [(500,000 - 100,000) / (1 + 0.10)Ai] + [(600,000 - 120,000) / (1 + O.IO)A2] + [(700,000 - 140,000) / (1 + O.IO)A3] - 1,000,000

[0091] NPV = [400,000 / 1.10] + [480,000 / 1.21] + [560,000 / 1.331] - 1,000,000

[0092] NPV » $272,727 + $396,694 + $420,771 - $1,000,000

[0093] NPV » $90,192

[0094] In this example, the NPV of the oil production project is approximately $90,192, which means the project is expected to generate a positive net value after accounting for the initial investment and future cash flows, assuming a 10% discount rate.

[0095] The metamodel 32 and the economic model 34 combined together can be considered as a forward model (depicted as 36 in FIG. 3) of the optimization problem.

[0096] A performance function 37 is used to evaluate if the value 35 obtained from the model 36 can be accepted or not, for example, by comparing with a target value. If the value 35 is accepted, it is output as value 38 for the entire optimization process. If the value 35 is not accepted, the solution needs further improvement and module 39 will be invoked.

[0097] Module 39 is configured to identify which optimization parameters are to be updated and by how much amount. This module can vary depending on the choice of optimization method used in the framework. For example, gradient-based methods compute gradients of loss with respect to the optimization parameters and use it to calculate an amount of update to be performed to each optimization parameter. Alternatively, metaheuristic methods such as particle swarm optimization, use heuristic approaches to calculate the pattern of updating.

[0098] Module 39 outputs a set of updated parameters, denoted as data 10, and feds back to the forward model 36. The above-described steps will be repeated until the performance function decides that the value 35 obtained can be accepted.

[0099] In some implementations, lower and upper bounds for one or more field control parameters may be defined and the module 39 may be configured to update a parameter within a boundary defined by the lower and upper bounds.

[0100] In the conventional optimization, full-scale reservoir simulations are run as part of a forward modelling process. However, due to large computational overhead, only a few updates can be performed, and the solution is not expected to improve much. The techniques disclosed herein provides a data-driven forward model that can allow fast forward computations and facilitates running a more rigorous optimization problem with very limited computational load. Those techniques canprovide a significantly better field development plan compared to the base input plan, thereby improving the business performance by a large margin. Moreover, those techniques can permit to perform optimization of the field-development plan in a computational inexpensive manner, thereby saving a significant amount of computational time that would be spent when seeking a good solution using the trial-and-error method on the actual reservoir simulations.

[0101] Since the forward model of the optimization method is data-driven, it is not a perfect replacement for the reservoir simulators. Rather, there would some error in the model predictions and this can affect the optimized results. To circumvent this, there are metrics such as R2-score, mean-squared error and mean-absolute error, among others, and these measures how far are the metamodel predictions from the ground-truth. However, from the perspective of optimization, the ranking of the simulations within the ensemble is also important. For example, the best simulation in terms of target objective value (e.g.) in the ground-truth set of simulations should also be ranked at the top by the metamodel. Failing to do so will lead to erroneous results from the side of the optimizer.

[0102] For the purpose of evaluating the ranking match between the ground-truth and the metamodel outputs, the following three different metrics can be considered.

[0103] Top-K AccuracyThis metric indicates how many of the top K simulations (in terms of any chosen economic parameters) from the ground-truth are also contained in the top K of the predictions generated by our trained metamodel. In other works, it looks at the intersection of the top-K solutions from the ground-truth and predictions and estimates the overlap.

[0104] Forward Top-K RobustnessIt is a robustness measure which checks if the top-ranking simulation from the ground-truth simulations is contained in the top K predictions.

[0105] Backward Top-K RobustnessIt is a robustness measure which checks if the top-ranking prediction obtained from the metamodel is contained in the top K ranking ground-truth simulations.[oio6] FIG. 4 shows an example of opportunity map generated based on results from a well-level metamodel according to aspects of the disclosure. As disclosed herein, the well-level metamodel is configured to determine a well KPI. The opportunity map is constructed to show the well KPI in different geological areas. As shown in Fig. 4, different KPI level can be shown with a colorful scalar. In the example shown in Fig. 4, the darker the area, the lower the KPI.

[0107] FIG. 5 shows an example of enhanced opportunity map according to aspects of the disclosure. As disclosed herein, physical constraints are incorporated in the opportunity map to mark areas where it should not be / cannot be drilled. As shown in the example of FIG. 5, the areas which should not be / cannot be drilled are marked as black.

[0108] FIG. 6 shows an example of opportunity map added with optimal well placement locations according to aspects of the disclosure. The areas with higher KPI (e.g., marked with lighter areas) generally indicates optimized locations for wells. A set of optimized locations for wells can be chosen and marked on the opportunity map.

[0109] While specific examples have been described herein, it will be obvious to those skilled in the art that various changes and modifications may be aimed to in the specification. It will, therefore, be understood by those skilled in the art that the particular embodiments of the invention presented here are by way of illustration only and are not meant to be in any way restrictive; therefore, numerous changes and modifications may be made, and the full use of equivalents resorted to, without departing from the scope of the invention.

Claims

Claims1. Computer-implemented method for determining well placement in a reservoir, the method comprising the steps of: a) performing a first process using at least one field-level metamodel to select field control and well configuration parameters, the field-level metamodel being adapted to mimic field behavior; b) performing a second process using at least one well-level metamodel to predict a well key performance indicator (KPI), the well-level metamodel taking into account downhole completion data and preferably well-level production and / or injection data, and field control and well configuration parameters selected from the first process; c) generating an opportunity map based on results from the second process; d) enhancing the opportunity map by incorporating physical constraints on well drilling locations; and e) determining well placement based on the enhanced opportunity map.

2. The method of the preceding claim, wherein the physical constraints on well drilling locations include geological features.

3. The method of the preceding claim, wherein the geological features comprise faults, existing well locations, or a combination thereof.

4. The method of any one of the preceding claims, wherein the first process is adapted to output an economic KPI, the economic KPI including a net present value (NPV) and / or ultimate total recovery (UTC).

5. The method of any one of the preceding claims, wherein the field-level metamodel is based on Powell method and / or Particle Swarm Optimization method.

6. The method of any one of the preceding claims, wherein the opportunity map represents a whole reservoir or a zone in the reservoir.

7. The method of any one of the preceding claims, wherein the opportunity map is generated for multiple timesteps.

8. The method of the preceding claim, wherein the opportunity map is updated in a subsequent timestep taking into account of areas for drilling selected in a previous timestep.

9. The method of any one of the preceding claims, further comprising a step of training the field-level metamodel and / or well -level metamodel based on simulation results in association with a plurality of simulation cases each representing a respective field control and well configuration scenario.

10. The method of the preceding claim, wherein the field control parameters include parameters relating to production facilities and parameters relating to well management, and wherein the well configuration parameters include parameters relating to well spacing, well pattern, a number of production and / or injection wells.

11. The method of any one of the preceding claims, further comprising generating a number of simulation cases with various field control and well configuration scenarios for training the field-level metamodel.

12. The method of any one of the preceding claims, wherein the step of selecting field control and well configuration parameters comprises performing a forward optimization process to obtain a set of field control and well configuration parameters.13- A system for determining well placement in a reservoir, the system comprising: a) means for performing a first process using at least one field-level metamodel to select field control and well configuration parameters, the field-level metamodel being adapted to mimic field behavior; b) means for performing a second process using at least one well-level metamodel to predict a well key performance indicator (KPI), the well-level metamodel taking into account downhole completion data and preferably well-level production and / or injection data, and field control and well configuration parameters selected from the first process; c) means for generating an opportunity map based on results from the second process; d) means for enhancing the opportunity map by incorporating physical constraints on well drilling locations; and e) means for determining well placement based on the enhanced opportunity map.

14. The system of claim 13, further comprising means for carrying out the method of any of claims 1 to 12.

15. A computer program comprising instructions which, when the program is executed by a computer, cause the computer to carry out the method of any of claims 1

Images