New patent application published for: ARTIFICIAL INTELLIGENCE FOR WELL PLACEMENT OPTIMIZATION

ARTIFICIAL INTELLIGENCE FOR WELL PLACEMENT OPTIMIZATION

Claims

1. An apparatus for computing, comprising: an input interface configured to receive: a base reservoir simulation representative of a geological oil reservoir; a processor configured to: cluster the reservoir into a plurality of sub-regions; generate, in a first optimization process, a candidate pool of well positions in the reservoir by screening and pruning a search space of possible well trajectories, each candidate well position being geologically plausible and satisfying one or more operational constraints; and optimize, in a second optimization process, the candidate pool of well positions to determine an optimized pool of wells that maximizes a dynamic variable of the reservoir; and an output interface configured to output the optimized pool of wells to a user.
2. The apparatus for computing of claim 1, wherein at least one of: the optimized pool of well positions specifies a set of producer wells and injector wells that is a subset of the candidate pool of well positions generated in the first optimization process; the base reservoir simulation comprises a set of cases with different maps of porosity and permeability that characterize geological uncertainty of the reservoir; and the processor is further configured, prior to clustering the reservoir, to perform a smart sampling sub-process on a plurality of reservoir realizations associated with geological uncertainty.
3. The apparatus for computing of claim 4, wherein: the processor is further configured, prior to clustering the reservoir, to perform a smart sampling sub-process on a plurality of reservoir realizations associated with geological uncertainty, and the smart sampling sub-process outputs a subset of K realizations selected from an original set of N realizations based on extracted feature representations of the realizations and clustering of the realizations in a projected feature space, the subset providing probability distributions similar to those of the original set of N realizations.
4. The apparatus for computing of claim 3, wherein the smart sampling sub-process determines the smallest subset of realizations that provides a similar expected value of a reservoir objective function under geological uncertainty as compared to the full set of realizations.
5. The apparatus for computing of claim 1, wherein the processor is further configured to post-process the output of K-means clustering to remove salt-and-pepper noise.
6. The apparatus for computing of claim 1, wherein the operational constraints include at least one of: a distance between two wells; a maximum number of injectors to drill; a maximum number of producers to drill; a distance between two producer wells; a distance between two injector wells; and a distance between a producer well and an injector well.
7. The apparatus for computing of claim 1, wherein, in the second optimization process, the processor is further configured to apply, within a derivative-free optimization loop, a projection operator to a candidate solution that is infeasible with respect to the operational constraints by replacing the candidate solution with a closest feasible well configuration that satisfies the operational constraints.
8. The apparatus for computing of claim 1, wherein the processor is further configured, as part of the second optimization process, to: train a surrogate model DNN to forecast a production response of the reservoir based on completion-related information, petrophysical information, reservoir cell-coordinate completion data, and time-varying well-control input; and use the trained surrogate model to optimize the pool of wells, wherein the surrogate model includes a spatial encoder configured to encode relationships among wells and/or well completions using the completion-related information and the petrophysical information to generate a spatial representation, a temporal encoder configured to process the time-varying well-control input conditioned on the spatial representation or a derivative thereof, a relevance-weighting mechanism configured to emphasize information most relevant to a predicted production response, and a prediction module configured to output a production forecast.
9. The apparatus for computing of claim 8, wherein the graph-based spatial encoder comprises a first graph-convolution layer configured to encode reservoir cell-coordinate completion data to generate a completion embedding, and wherein the completion embedding is concatenated with the petrophysical information before processing by the one or more additional graph-convolution layers.
10. The apparatus for computing of claim 8, wherein the one or more additional graph-convolution layers are configured to output a pooled spatial representation by a global add pooling operation.
11. The apparatus for computing of claim 10, wherein the pooled spatial representation is used to initialize a hidden state of the temporal sequence encoder.
12. The apparatus for computing of claim 8, wherein the temporal encoder comprises a recurrent architecture selected from a gated recurrent unit (GRU) layer and a long short-term memory (LSTM) layer, or a temporal convolution module configured to receive a sequence of the time-varying well-control input.
13. The apparatus for computing of claim 12, wherein the relevance-weighting mechanism comprises an attention mechanism or another feature-selection or gating mechanism configured to emphasize information most relevant to the predicted production response, and wherein the prediction module is configured to output a time-series forecast of oil production, gas production, water production, injection response, or a combination thereof, at a well level, a field level, or both.
14. A method of optimizing a set of oil-producing wells to be drilled in a reservoir, comprising: receiving a base reservoir simulation representative of a geological oil reservoir; clustering the reservoir into a plurality of sub-regions; generating, in a first optimization process, a candidate pool of well positions in the reservoir by screening and pruning a search space of possible well trajectories, each candidate well position being geologically plausible and satisfying one or more operational constraints; and optimizing, in a second optimization process, the candidate pool of well positions to determine an optimized pool of wells that maximizes a dynamic variable of the reservoir; and outputting the optimized pool of wells to a user.
15. The method of claim 14, further comprising, prior to clustering the reservoir, performing a smart sampling sub-process on a plurality of reservoir realizations associated with geological uncertainty.
16. The method of claim 15, wherein the smart sampling sub-process outputs a subset of K realizations selected from an original set of N realizations based on extracted feature representations of the realizations and clustering of the realizations in a projected feature space, the subset providing probability distributions similar to those of the original set of N realizations.
17. The method of claim 16, wherein the smart sampling sub-process determines the smallest subset of realizations that provides a similar expected value of a reservoir objective function under geological uncertainty as compared to the full set of realizations.
18. The method of claim 14, wherein at least one of: the clustering includes computing different clusters of the reservoir, each cluster representing a sub-region of the reservoir, using a clustering algorithm, or the operational constraints include at least one of: a distance between two wells; a maximum number of injectors to drill; a maximum number of producers to drill; a distance between two producer wells; a distance between two injector wells; and a distance between a producer well and an injector well.
19. The method of claim 14, further comprising, as part of the second optimization process: training a DNN surrogate model to forecast a production response of the reservoir based on completion-related information, petrophysical information, reservoir cell-coordinate completion data, and time-varying well-control input; and using the trained DNN surrogate model to optimize the pool of wells, wherein the surrogate model includes a spatial encoder configured to encode relationships among wells and/or well completions using the completion-related information and the petrophysical information to generate a spatial representation, a temporal encoder configured to process the time-varying well-control input conditioned on the spatial representation or a derivative thereof, a relevance-weighting mechanism configured to emphasize information most relevant to a predicted production response, and a prediction module configured to output a production forecast.
20. The method of claim 19, wherein, at least one of: the graph-based spatial encoder comprises a first graph-convolution layer configured to encode reservoir cell-coordinate completion data to generate a completion embedding, and the completion embedding is concatenated with the petrophysical information before processing by the one or more additional graph-convolution layers, or the one or more additional graph-convolution layers output a pooled spatial representation by a global add pooling operation.