Enhancing Natural Gas Production Prediction Using Machine Learning Techniques: A Study with Random Forest and Artificial Neural Network Models

Abstract

Abstract The International Energy Agency (IEA) in 2020 predicted that the supply of natural gas will expand globally by 4.1 percent and would continue to grow as a result of rising demand for energy in the industrial and power sectors. In order to ensure well integrity, prepare for potential interventions, and improve the energy mix and transition, accurate gas production prediction is essential. Various models exist for predicting Natural Gas production, but currently used tools lean towards machine learning. However, Machine Learning models in extant literature yield inaccurate results due to limitations of data, predictive range, and robustness. To address this issue, this study develops a model that combines traditional and non-traditional features to reduce computational time and increase robustness. The model uses two algorithms, Artificial Neural Network (ANN) and Random Forest (RF) Regression, and over 13,500 data points from 15 years of production. To assess the performance of this model, four statistical error metrics were employed; mean squared error (MSE), root mean squared error (RMSE), coefficient of determination (R2) and correlation coefficient (R). The ANN model achieved an accuracy of 0.93 with a MSE of 0.1244, a validation score of 0.1171, and a testing score of 0.1337. The RF Model achieved accuracy of 0.96 and an R of 0.98. Feature selection was done on the model to provide insights on how each feature affects the model accuracy. With flowing tubing hole pressure produced contributing over 50 percent to the predictive capacity and shut-in tubing hole pressure with less than 5 percent. In comparison with existing models, the developed models performed better in terms of predictive accuracy, robustness, predictive data range and computational time. The characteristics of the developed model for which novelty is claimed include; the explicit nature of the model that helps it be deployed in software application, the low computational time and its high predictive capacity. The model created for this study will aid in forecasting and predicting gas production, guaranteeing its efficacy, economy, and robustness. It also serves as a starting point for further investigation into natural gas forecasting and prediction.