Automated Pressure Distribution in Hydraulically Fractured Reservoirs by Machine Learning Algorithms

Abstract This paper introduces an approach for establishing a pressure map of unconventional reservoirs that are developed by multiple hydraulic fractures. The approach focuses on utilizing machine learning algorithms for simulating the pressure distribution in the three porous media of these reservoirs (Hydraulic fractures, stimulated reservoir volume (SRV), and unstimulated reservoir volume (USRV)). The objective is developing an integrated protocol that could be used to predict the pressure behavior of different reservoirs knowing their geometrical configurations and petrophysical properties of the porous media. The methodology, followed in this study, includes several tasks. The first is estimating the pressure distribution numerically by solving the analytical pressure drop models of these reservoirs at different production times considering different reservoir and fluid characteristics. Purposefully, the pressure drop is calculated at different points (locations) in the three porous media. The second is modifying three machine-learning models for the pressure distribution of the three porous media. These models consider the flow regimes that could be occurred due to the flow in these porous media (Hydraulic fracture, bi-linear, and formation linear flow regimes). The three models are developed and trained using the numerically-calculated pressure data. The modified machine-learning models, in the third task, are used to predict the pressure map of the hydraulically fractured reservoirs wherein the geometrical configuration and petrophysical properties are known. In the fourth task, the proposed models are tuned and validated using real pressure data to examine the reliability of these models. The study has reached several conclusions regarding the pressure distribution in the three porous media of fractured reservoirs. The pressure distribution along the hydraulic fractures, SRV, and USRV demonstrate an exponential behavior with the distance during the entire production time. The automated pressure distribution using the ensemble Kalman filter and the random forest algorithms confirms the exponential behavior of the pressure distribution. Identifying the flow regimes by the random frost is a keystone in predicting the pressure distribution. The most influential parameters on the pressure distribution are determined by the random frost. They are the hydraulic fracture conductivity, the fracture/SRV hydraulic ratio, and the SRV/USRV hydraulic ratio. The analytical solutions can be used to develop machine learning-based models; however, tuning is required for the utilization of these models to analyze the real data. This study introduces a novel approach for predicting the pressure distribution (pressure map) of hydraulically fractured reservoirs using the machine learning technique. The data-driven models by the machine learning approach can be confidently used to predict the pressure distribution with no need to have the pressure records if the most influential parameters are available.