Artificial Intelligence Applications to Forecast Oil Production from Hydraulically Fractured Reservoirs

Applications of artificial intelligence (AI) are rapidly growing in every field from medical diagnosis to robot control due to this method’s innovative techniques in connecting inputs to outputs and recent advancements in high-speed computing. The oil and gas industry has started exploring the possibilities of AI applications in various sectors such as reservoir characterization, fluid thermodynamic properties, production maximization, fracture job optimization, recoverable hydrocarbon estimation, well placement using pattern recognition, etc. In this study, two popular artificial intelligence techniques, namely the Least Square Support Vector Machine (LSSVM) and Artificial Neural Networks (ANN) methods are developed as surrogate models to forecast oil production from hydraulically fractured reservoirs such as shales and other tight formations. The objective of this work is to evaluate AI techniques as alternative methods for production performance analysis. Eight important factors for production such as reservoir permeability, gas relative permeability exponent, rock compressibility, initial gas oil ratio, slope of gas oil ratio versus pressure, initial pressure, flowing bottom hole pressure and fracture spacing are selected as input for this study. The ranges of inputs are based on real field data from the Eagle Ford in Texas, Bakken in North Dakota and Niobrara in Colorado-Wyoming in the USA. Two important production performance measures, i.e. oil recovery and produced gas oil ratio (GOR) are selected as output of the models. Oil recovery and gas recovery are recorded at particular time intervals (90 days, 1 year, 5 years, 10 years, and 15 years) and after the oil rate drops to 5 barrels per day per fracture. Therefore, five time-based models and one rate-based model are developed using LSSVM and ANN for oil recovery and GOR. Surrogate models (LSSVM and ANN) are trained with 80% of data (training set) and are verified with 20% of data (test set) based on numerical simulation. Particle Swarm Optimization (PSO) is used to find the model parameters. Fitness of the models are evaluated by measuring the coefficient of determination and normalized root mean square error. LSSVM models have shown very high accuracy compared to ANN models. The complex nature of produced GOR are successfully predicted by LSSVM.  These models can be used as proxy models for reservoir simulator to forecast recovery and to perform sensitivity and uncertainty analyses.