Date of Award

Fall 2017

Document Type


Degree Name

Doctor of Philosophy (PhD)


Electrical and Computer Engineering

First Advisor

Povinelli, Richard

Second Advisor

Nourzad, Farrokh

Third Advisor

Corliss, George


This dissertation focuses on quantifying forecast uncertainties in the energy domain, especially for the electricity and natural gas industry. Accurate forecasts help the energy industry minimize their production costs. However, inaccurate weather forecasts, unusual human behavior, sudden changes in economic conditions, unpredictable availability of renewable sources (wind and solar), etc., represent uncertainties in the energy demand-supply chain. In the current smart grid era, total electricity demand from non-renewable sources influences by the uncertainty of the renewable sources. Thus, quantifying forecast uncertainty has become important to improve the quality of forecasts and decision making. In the natural gas industry, the task of the gas controllers is to guide the hourly natural gas flow in such a way that it remains within a certain daily maximum and minimum flow limits to avoid penalties. Due to inherent uncertainties in the natural gas forecasts, setting such maximum and minimum flow limits a day or more in advance is difficult. Probabilistic forecasts (cumulative distribution functions), which quantify forecast uncertainty, are a useful tool to guide gas controllers to make such tough decisions. Three methods (parametric, semi-parametric, and non-parametric) are presented in this dissertation to generate 168-hour horizon probabilistic forecasts for two real utilities (electricity and natural gas) in the US. Probabilistic forecasting is used as a tool to solve a real-life problem in the natural gas industry. A benchmark was created based on the existing solution, which assumes forecast error is normal. Two new probabilistic forecasting methods are implemented in this work without the normality assumption. There is no single popular evaluation technique available to assess probabilistic forecasts, which is one reason for people’s lack of interest in using probabilistic forecasts. Existing scoring rules are complicated, dataset dependent, and provide less emphasis on reliability (empirical distribution matches with observed distribution) than sharpness (the smallest distance between any two quantiles of a CDF). A graphical way to evaluate probabilistic forecasts along with two new scoring rules are offered in this work. The non-parametric and semi-parametric probabilistic forecasting methods outperformed the benchmark method during unusual days (difficult days to forecast) as well as on other days.