Horizontal well flow rate prediction applying machine-learning model

Abstract

Relevance. The need to accurately and quickly predict flow rates of horizontal wells. This allows optimizing drilling schedules, enhanced oil recovery programs, and field development strategy, as well as making the economic model more accurate and predictable. Currently, analytical calculations and numerical modeling methods are used to predict well production rates. These methods have limitations in both accuracy and time. To solve this problem, it is proposed to use machine learning, which due to its accuracy, adaptability, and speed, allows excluding the disadvantages of the above-mentioned methods. Aim. To create a machine-learning model to quantify gas well flow rates based on geological properties at different time steps. Object. Stock of horizontal wells in a gas condensate field in Western Siberia. Methods. Mathematical modelling, machine learning and statistical methods. Results. The authors have carried out 300 iterations of hydrodynamic modeling in a simulator. They collected an initial data set with the following parameters: time step, porosity, permeability, initial water saturation, reservoir thickness, bottom hole pressure at different distances from the wellbore, and gas flow rate. Machine learning models based on random forest and gradient boosting algorithms were created with different ratios of testing/training samples. The machine learning models were able to accurately predict the gas flow rate of a horizontal well. Gradient boosting showed better prediction results compared to random forest: root mean square error is equal to 8440 std. m3/day, mean absolute percentage error is equal to 3,95%, and coefficient of determination (R2)=0,991.

Keywords

• machine learning
• gradient boosting
• random forest
• horizontal well flow rate prediction
• Darcy's law
• reservoir simulation