ADAPT seminar

Speaker: Dr. Eugene Morgan (Assistant Professor of Petroleum and Natural Gas Engineering, PSU)
Topic: "Big Data in Oil and Gas: An Application of Time Series Analysis to Improve Probabilistic Forecasting of Production from Marcellus Wells"

Room: 529 Walker Building (refreshments served)
Time: Thursday September 07, 2017 1:30pm - 2:30pm
Abstract:
As the number of oil and gas wells, and the amount of data collected from them, continues to proliferate, the energy industry shifts more towards data analytics to inform decision making. There exists a research opportunity here to learn more about reservoir behavior, and enhance our predictive capabilities for exploration and production purposes, through Big Data analytics. The creation of a “virtual asset”, which brings together all forms of data from public and private sources for one oil or gas field into one accessible source, would facilitate these sorts of analyses, as well as provide a resource for innovating undergraduate and graduate education.

An example application is demonstrated upon an incipient “virtual asset”, whereby probabilistic forecasts of gas production from Marcellus wells are improved through the use of time series statistics. Traditionally, such forecasting is achieved empirically through the fitting of preconceived, non-linear models called “decline curves”. As the name implies, these model the decay in gas production rate with time since initial production. However, these decline curves often fail to capture the full behavior of production from shale gas reservoirs, such as the Marcellus formation. The residuals from fitting these decline curves to historical production data often exhibit significant serial autocorrelation. Including an autoregressive term with the decline model improves the goodness-of-fit, the predictive accuracy, and the prediction interval, as demonstrated through Bayesian regression on 1,007 Marcellus wells. In conclusion, we find that rather than formulating new decline models, the behavior of shale gas production can be better captured through statistical advances, even basic ones.