Off-campus Michigan Tech users: To download campus access theses or dissertations, please use the following button to log in with your Michigan Tech ID and password: log in to proxy server

Non-Michigan Tech users: Please talk to your librarian about requesting this thesis or dissertation through interlibrary loan.

Date of Award


Document Type

Campus Access Dissertation

Degree Name

Doctor of Philosophy in Computer Science (PhD)

Administrative Home Department

Department of Computer Science

Advisor 1

Laura E. Brown

Committee Member 1

Nilufer Onder

Committee Member 2

Timothy Havens

Committee Member 3

Qiuying Sha


To guarantee sufficient electricity supply for its highest demands, many regional organizations surcharge their customers during coincident peaks (CPs), a time of highest demand across the system or region of interest. Therefore, the accurate prediction of these coincident peaks would be helpful not only for companies to ensure sufficient generation is available, but also for customers who may try to avoid electricity consumption and consequent additional cost. This dissertation focuses on the prediction of the top five coincident peak days (5CPs) in a year. We use classification models to solve this imbalanced prediction problem (around 1.4% for positive cases) by classifying the next day as 5CP days or non-5CP days.

We analyze six sets of actual historical data from different regions of Canada and the United States. We explore the effect of forecast accuracy on 5CP days prediction through four cases: I - knowing tomorrow’s power demand and weather condition exactly (an oracle), II & III - knowing some information about tomorrow (an oracle + increasing noise), and IV - no knowledge of future. We propose a three-phase model to predict 5CP days: first, clustering is applied to filter some negative cases, second, an all convolutional neural network that estimates the probability of being a 5CP day for the remaining cases is learned, and third, an adaptive method is used determines thresholds. This three-phase model exhibits promising performances with the highest mean recall of 1.00, mean precision of 0.56, and mean F1 score of 0.72. Finally, we explore the use of a few-short learning framework for this problem. A triplet network is implemented for the 2-way-5-shot classifications. The prediction results have the highest mean recall of 1.00, mean precision of 0.67, and mean F1 score of 0.79.