Date of Award
Fall 2016
Document Type
Thesis
Degree Name
Master of Science (MS)
Department
Electrical and Computer Engineering
First Advisor
Brown, Ronald H.
Second Advisor
Corliss, George F.
Third Advisor
Johnson, Michael T.
Abstract
This thesis explores techniques by which the accuracy of gas demand forecasts can be improved during extreme cold events. Extreme cold events in natural gas demand data are associated with large forecast error, which represents high business risk to gas distribution utilities. This work begins by showing patterns associated with extreme cold events observed in natural gas demand data. We present a temporal pattern identification algorithm that identifies extreme cold events in the data. Using a combination of phase space reconstruction and a nearest neighbor classifier, we identify events with dynamics similar to those of an observed extreme event. Results obtained show that our identification algorithm (RPS-kNN) is able to successfully identify extreme cold events in natural gas demand data. Upon identifying the extreme cold events in the data, we attempt to learn the residuals of the gas demand forecast estimated by a base-line model during extreme cold events. The base-line model overforecasts days before and underforecasts days after the coldest day in an extreme cold event due to an unusual response in gas demand to extreme low temperatures. We present an adjustment model architecture that learns the pattern of the forecast residuals and predicts future values of the residuals. The forecasted residuals are used to adjust the initial base model’s estimate to derive a new estimate of the daily gas demand. Results show that the adjustment model only improves the forecast in some instances. Next, we present another technique to improve the accuracy of gas demand forecast during extreme cold events. We begin by introducing the Prior Day Weather Sensitivity (PDWS), an indicator that quantifies the impact of prior day temperature on daily gas demand. By investigating the complex relationship between prior day temperature and daily gas demand, we derived a PDWS function that suggests PDWS varies by temperature and temperature changes. We show that by accounting for this PDWS function in a gas demand model, we obtain a gas model with better predictive power. We present results that show improved accuracy for most unusual day types.