Flowing Bottomhole Pressure Prediction with Machine Learning Algorithms for Horizontal Wells
Sohrat Baki, Serkan Dursun
Abstract
Abstract Flowing Bottom Hole Pressure (FBHP) is one of the critical parameters during evaluation of unconventional oil and gas horizontal wells. The FBHP is mainly used for production optimization, calculation of productivity index and assessment of well performance evaluation with respect to offset wells. To calculate FBHP, production engineers use physics-based models and empirical correlations calibrated with actual physical measurements from offset wells. The objective of the paper is to develop a data driven approach for prediction of the FBHP with machine learning algorithms during post frac flowback period with multi-phase flow regimes. Surface production data from 30+ wells were collected with the key attributes such as wellhead pressure, gas rate, condensate rate, water rate, wellhead temperature, choke size, fluid density. FBHP was calculated based on physics-based software calibrated with physical downhole pressure gauges. In machine learning workflow, data collection and cleaning, exploratory data analysis, train-test split, pre-processing, cross-validation and tuning tasks were employed to obtain the accurate and generalized predictive models. Ensemble tree-based, support vector machines, and neural network models were explored in this study and their performances were evaluated by several regression error metrics. Analysis of model results showed the potential of machine learning approach for predicting the FBHP especially in multiphase flow regime. The model results showed accurate prediction of the FBHP within acceptable confidence interval application period. It is important to note that model accuracy dependents on testing period and flow condition as expected from interpolation nature of machine learning predictive models. The models were applied to new wells which were not used in the training process. The prediction results on the new wells showed outstanding performance and thus enabled decision making for wide area of application. This study shows that machine learning based predictive models can learn the patterns in complex parameters governing the FBHP behavior. Furthermore, robust machine learning models can be used for different phase of production period to optimize field development.