Electronic Thesis and Dissertation Repository

Thesis Format

Monograph

Degree

Master of Engineering Science

Program

Electrical and Computer Engineering

Collaborative Specialization

Artificial Intelligence

Supervisor

Capretz, Miriam A.M.

2nd Supervisor

Sadhu, Ayan

Co-Supervisor

Abstract

Peak demand consumption is an ongoing research topic due to current environmental concerns. Accurate prediction of peak demand leads to improved power storage scheduling and smart grid management. However, existing researches on peak demand in commercial buildings lack focus on the timestamps for the incident of peak consumption. For this reason, this research proposes to label three indexes per day and use the novel Energy Peaks and Timestamping Prediction (EPTP) framework to detect the energy peaks as well as the timestamps for the occurring indexes.

The EPTP framework is proposed with three phases. In the first phase, data preprocessing cleans the raw data into the intended input for the deep learning model, and timestamp labelling creates the expected output for training and evaluation of the model. The second phase focuses on energy consumption prediction using Long Short-Term Memory (LSTM) network, which is dedicated to processing sequential data. The last phase uses Multilayer Perceptron (MLP) for the purpose of timestamp prediction. The EPTP framework is evaluated using various data resolutions and compared to the common label of using block maxima from extreme value theory. Specifically, the two-hour hit rate improves from 21\% using the block maxima approach to 52.6\% with the proposed EPTP framework, and from 65.3\% to 86\% for the 1-hour resolution and the 15-minute resolution, respectively. In addition, the average minute deviation decreases from 120 minutes using the block maxima approach to 62 minutes with the proposed EPTP framework for the high-resolution data. The framework shows adequate results from high-resolution data using real-world commercial supermarket energy consumption.

Summary for Lay Audience

Building energy consumption takes up more than 30\% of total global energy use and accounts for 27\% of total greenhouse gas emissions. Peak demand management serves the purpose of effective building energy use, which in turn decreases carbon footprint. Accurate peak demand in commercial buildings benefits the suppliers by improving efficiency in electricity production; and the consumers by reducing fluctuations and energy waste. Accurately predicting energy peaks can lead to proactive peak-shaving strategies and battery response schedulings. In the existing research, most of the effort has been dedicated to predicting the intensity or the amount of energy used during peak demand. However, there has been a lack of interest in extracting the timing of peak demand, especially in commercial buildings.

In this research, the effort is dedicated to not only predicting the energy consumption during the peak hours but also the index or the timestamps for the instance of which peak has occurred. A three-phase methodology for Energy Peaks and Timestamping Prediction (EPTP) is proposed in this work. Data preprocessing and timestamp labelling create the intended input for the deep learning model. Energy consumption prediction is performed with a Long Short-Term Memory (LSTM) network, which is dedicated to processing sequential data; followed by the timestamp prediction using Multilayer Perceptron Network. The proposed methodology is validated through experiments using real-world commercial supermarket data. Results show that the 2-hour hit rate reaches 86\% with 62 minutes deviation in the peaking index with 15-minute resolution data.

Creative Commons License

Creative Commons Attribution 4.0 License
This work is licensed under a Creative Commons Attribution 4.0 License.

Download

Share

COinS