Forecasting long-term world annual natural gas production by machine learning

Abstract

The goal of this study is to model the global annual natural gas production using a variety of machine learning models in order to predict future production and determine a peak production date. World gross domestic product (GDP) based on purchasing power parities (at PPPs), inflation percentage, Henry Hub Price, Eum Price, cumulative natural gas resources, and annually discovered new resources were taken as descriptor variables, and Shapley analysis was conducted to observe the importance of features on the dataset. It was revealed according to this analysis that, Henry Hub price, inflation percentage, and newly discovered resources had minor effects on natural gas production, so they were left out. Then, a variety of machine learning algorithms were employed and the one with the highest prediction ability was found to be the stochastic gradient descent (SGD) algorithm. Next, this model was tested under four different scenarios, each with different GDP and natural gas price projections. Finally, natural gas production was found to reach its peak sometime between 2034 and 2046. It was then concluded that rather than relying on a traditional approach based on the Hubbert Curve, a machine learning model that takes into account all relevant factors can be used to accurately forecast natural gas production and its peak time, allowing governments and policymakers to make the necessary preparations.