Machine Learning Assisted Well Placement Optimization

Abstract

Abstract Well placement optimization is a complicated problem which is usually solved by direct combination of reservoir simulators with optimization algorithm. However, depending on complexity of the reservoir model studied, thousands of simulations is usually needed for accurate and reliable results. In this research, we present a novel approach – machine learning (ML) assisted proxy model that combines reservoir simulations and reduced physics model to reduce computational cost. In the proposed model framework, first several (depending on the complexity of the problem) uniformly distributed random coordinates are selected. These chosen coordinates are considered as data points for ML model. For the chosen coordinates (training set) reservoir simulations are executed and NPV/recovery values are calculated (target variable). Spatial locations as well as petrophysical properties of the same coordinates extracted from simulation model are also used as an input to the ML model. ML model is further improved by combining with Fast Marching Model (FMM) which is a robust reduced physics model. The inclusion of FMM helps identify drainage volume for producers and hence enhance model training. Finally, the trained ML model is coupled with stochastic optimization algorithms to determine infill well location with highest NPV/recovery. Using an example field data, we present two specific cases of using proposed model: a) for greenfield with a single new well, b) for greenfield with multiple new wells. Results indicate that developed ML model can predict NPV with around 96% accuracy (testing score). This gives great confidence in predictions from the trained hybrid model that can be used as a proxy model for reservoir simulations. Coupling the trained hybrid model with Particle Swarm Optimization (PSO), the location of the new producers with maximum NPV are determined. The results are further confirmed with an exhaustive search from all potential locations. A novel approach is presented to show how traditional physics-based approaches can be combined with machine learning algorithms to optimize well placement. Such approaches can be integrated in current greenfield and brownfield reservoir engineering workflow to drastically reduce decision making times.