Analyzing the temperature dimension of University of Illinois electricity demand

Guerrero, Spencer White

Permalink

https://hdl.handle.net/2142/90067 Copy

Description

Title

Analyzing the temperature dimension of University of Illinois electricity demand

Author(s)

Guerrero, Spencer White

Issue Date

2015-07-22

Department of Study

Atmospheric Sciences

Discipline

Atmospheric Sciences

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

M.S.

Degree Level

Thesis

Keyword(s)

• singular spectrum analysis (SSA)
• locally weighted scatterplot smoothing (LOESS)
• Autoregressive Integrated Moving Average (ARIMA)
• electricity demand
• short term load forecasting
• clustering

Abstract

The University of Illinois Utilities Administration Office currently relies on outside consultants to model next day electricity demand. The forecasted demand is used to determine how much electricity to purchase and to lock in prices on the next day market. This study focuses on the development of an electricity demand forecast model using modern numerical weather prediction output to account for the temperature dependency in load forecasting. This model, hereafter referred to as the “dispatch model” used singular spectrum analysis to first de-trend the data, LOESS fits were used quantify the relationship between temperature and electrical demand. LOESS fits were dependent on campus school session dates as well as a Mon-Fri and Sat, Sun weekday/weekend split. Lastly an ARIMA model was used on the residuals of the various LOESS fits to predict short-term electricity demand. Beyond quantifying the temperature dependence in electricity demand, this study considered two buildings that are responsible for approximately 25% of campus electricity demand, the Blue Waters super computer and the Oak Street Chiller Plant. Analysis of the Blue Waters electricity consumption did not suggest a strong dependence on temperature. Given the importance of these two buildings, they were analyzed individually. To evaluate the impact of temperature forecasts, a naive forecast model as well as a seasonal naive model were back tested on the data, in order to establish a baseline for model accuracy. The dispatch model was back tested on the data using perfect temperature forecasts, in order to establish a maximum upper bound for model performance. Results suggest, if perfect temperature forecasts were available, that the dispatch model would outperform the benchmark models' 2 day or longer mean absolute error even for 30 day forecasts for electricity demand. Finally, the dispatch model was back tested using actual temperature forecasts. The performance of the dispatch model, using actual temperature forecasts, was compared to the benchmark models’ performance and results demonstrate that the dispatch model outperforms both benchmark models by reducing mean absolute percent error by 50%. In addition to the demonstration of more accurate modelling, results from this analysis on campus electricity demand were useful to Utilities Administration staff.

Graduation Semester

2015-08

Type of Resource

text

Permalink

http://hdl.handle.net/2142/90067

Copyright and License Information

Copyright 2015 Spencer Guerrero

Owning Collections