Date of Award
Master of Science (MS)
Mathematics, Statistics and Computer Science
This thesis examines alarm forecasting methods for a natural gas production pipeline to assure the efficient transportation of high-quality natural gas. Natural gas production companies use pipelines to transport natural gas from the extraction well to a distribution point. Forecasting natural gas pipeline pressure alarms helps control room operators maintain a functioning pipeline and avoid costly down time. As gas enters the pipeline and travels to the distribution point, it is expected that the gas meets certain specifications set in place by either state law or the customer receiving the gas. If the gas meets these standards and is accepted at the distribution point, the pipeline is referred to as being in a steady-state. If the gas does not meet these standards, the production company runs the risk of being shut-in, or being unable to flow any more gas through the distribution point until the poor-quality gas is removed.Sensors are used to collect real-time gas quality information from within the pipe, and alarms are used to alert the control operators when a threshold is exceeded. If operators fail to keep the pipeline’s gas quality within an acceptable range, the company risks being shut¬¬-in or rupturing the pipeline. Predicting gas quality alarms enables operators to act earlier to avoid being shut-in and is a form of predictive maintenance. We forecast alarms by using a 10th-order autoregressive model, autoregressive model with exogenous variable, simple exponential smoothing with drift (Theta Method) and an artificial neural network with alarm thresholds. The alarm thresholds are defined by the production company and are occasionally adjusted to meet current environment conditions. The results of the alarm forecasting method show that we accurately forecast natural gas pipeline alarms up to a 30-minute time horizon. This translates into sensitivity rates that drop from around 100% at one minute to 82.7% at a 30-minute forecast horizon. This means that at 30 minutes, we correctly forecast 82.7% of the alarms. All alarm forecasting models outperform the state-or-the-art forecaster used by the production company, with the artificial neural network performing the best.