Date of Award

Spring 2019

Document Type


Degree Name

Doctor of Philosophy (PhD)


Electrical and Computer Engineering

First Advisor

Brown, Ronald H.

Second Advisor

Corliss, George F.

Third Advisor

Povinelli, Richard J.


Local natural gas distribution companies rely on accurate forecasts of daily demand to buy gas and deliver it to their customers. To forecast consumption, mathematical models with inputs such as weather and historical daily demand are considered. Many needs exist in the energy industry where the frequency of measurement is different from demanded. When the needed forecast frequency is higher than the measurements, disaggregation approaches are needed. We built multi-parameter linear regression models using monthly data. Several decoration methods in the disaggregation process are developed to improve the model accuracy. Prior-day weather adjustment is used to capture the daily fluctuation of gas consumption as a result of the temperature differences between current day and prior day. Also, behavioral patterns in gas consumption are incorporated in the models to account for consumption patterns in weekdays vs. weekend and days of week. Furthermore, we consider long-term characteristics in the gas consumption data originated from population changes, differences in building efficiency, and economic impacts. Firstly, Base Load Trend and later Heat Load Trend are considered in the linear regression models. Secondly, historical flow is detrended to act like the most recent data by altering its characteristics to approximate a stationary customer base with current behavioral patterns. Root Mean Square Error, Mean Absolute Percent Error, and Weighted Mean Absolute Percent Error are used as means for assessing the performance of our approaches. All decorations enhance forecasts, with Prior-Day adjustment as the most effective. The combination of decorations leads to further enhancements. Inclusion of detrending models decreases the forecast errors significantly. For geographic areas that have experienced identifiable trends, considering Base Heat Load Trend in the model shows the most improvement in detrending models. Extensive comparisons between decoration and detrending algorithms and the combination of these models show all methods enhance daily gas demand forecast accuracies. The combination of Base Heat Load Trend model, Day of the Week, and Prior-Day adjustment is most effective to improve the accuracy of daily demand forecasts from historical monthly gas consumption without need to any additional infrastructure to save Local Distribution Companies and customers a large amount of money.

Included in

Engineering Commons