Production screening and optimization using smart proxy modeling

Numerical models are the primary tools to look into the fluid flow behavior in the complex and uncertain reservoir environment. Engineers use numerical models to perform crucial tasks in reservoir engineering, such as uncertainty quantification, history matching, production forecasting, and optimization, to eventually make the best decisions for field development. A conventional numerical model often consists of millions of grid blocks, and depending on the level of complexities within the model, it may take hours or days to perform a single run. A comprehensive study of a numerical reservoir model requires hundreds or thousands of repetitions, making the decision very costly and time-intensive. Proxy modeling is a solution for the computational cost related to the numerical models. They make a relationship between the input design parameters and the desired outputs by using various statistical/mathematical/data-driven underlying models. Nevertheless, they have their own limitations. The biggest disadvantage of the conventional proxy models is that they cannot keep the complexities within the reservoirs. It means they have no or limited sense of objects that exist in the reservoirs such as faults, boundaries, wells, etc. The main objective of this research is to present smart proxy modeling (SPM) as a substitute for numerical models to address the computationally expensive and time-consuming drawbacks of numerical models and find a solution for keeping the complexities within the reservoir as conventional proxy models have. SPM is developed based on the implementation of pattern recognition and machine learning techniques, and it has an additional feature engineering step compared to the traditional known proxy models in the literature. The feature engineering step extracts new static and dynamic parameters from the numerical model. The constructed SPM takes only a few seconds to perform a single run. The SPM in this research is developed in grid- based and well-based types. The grid-based SPM can predict the grids’ properties, such as fluid saturation and pressure, and the well-based SPM is used to predict well production. Furthermore, the parallel implementation of the well-based SPM with grid-based SPM (named hybrid well-based SPM) is tested in this research. The proposed SPM in this research is modified at different construction steps compared to existing SPMs in the literature that suffer from construction efficiency and reliability. Based on our literature review, we target our investigation into techniques to improve efficiency and accuracy by focusing on sampling, feature ranking, and underlying model construction. In existing SPM literature, only one technique is used during each construction step where there are opportunities to explore novel construction steps to improve overall SPM accuracy and efficiency. The presented sequential sampling technique avoids repeating the construction procedure from resampling and running the high-fidelity model, thereby saving time and making the SPM workflow more efficient. In the feature ranking step, an average of multiple ranking algorithms is used to find the best subset of input parameters which eventually helps the overall efficiency in the feature selection step. The performance of the convolutional neural network (CNN) as the underlying model is also tested and compared to the implemented artificial neural networks (ANN) in the literature. In this research, the SPMs are constructed for two case studies. The first case study corresponds to a waterflooding scenario for the offshore Norway Volve field. The design parameters involve five parameters of the wells’ liquid production rates, and the objectives are to screen and optimize oil recovery. For the screening purpose, the grids’ pressure and oil saturation are considered as the outputs of the grid-based SPM. For production optimization, the wells' cumulative oil production is the output of the well-based SPM. Finally, the performance of well- based SPM coupled with two derivative-free optimizers, particle swarm optimization and genetic algorithm, are compared. The SPM with ANN underlying model provides an accuracy of 89-92% compared to the 94-99% of the CNN technique for the grid-based SPMs. However, for the well-based SPM, the goodness of fit for the 1D-CNN model is similar to the ANN model, but its accuracy (presented in MAPE) is slightly better than ANN. The well-based for this case study is coupled with PSO and GA optimization algorithms to find the best selection of designing parameters (individual well’s LPR) and to maximize the cumulative oil production over ten years. Both optimizers are quite successful in finding the global optimum. Nevertheless, PSO shows a more reliable and faster convergence to the solution. The second case study corresponds to a water alternating gas (WAG) scenario for the offshore Norway Norne field. This case study aims to test the whole procedure of SPM construction in another field with different levels of complexities and more design parameters. The design parameters for the WAG scenario are nine parameters of gas/water injection cycle, field gas/water injection rate, gas/water injection distribution between two injectors, and injectors’ BHPs. Similar to the first case study, screening and oil recovery optimization are the targets for this case study. The trained CNN models give an accuracy of 85-87% for different timesteps of the grid-based dataset at the blind test. However, after adding five more sample points using the sequential LHS, the accuracy increases to 94-99%. The well-based SPM, similar to the first case study, does not give promising improvement in terms of accuracy