Using Machine Learning Methods Towards Identifying College Campus Load Profiles and Energy Storage Application for Reducing Peak Energy Demand From the Utility Grid

Abstract

Efforts to reduce peak energy demand on the utility grid have been a challenge due to unique load profiles for individual customers such as college campuses, businesses, and homeowners. This work illustrates the application of machine learning in the form of Bayes Estimation, Principal Component Analysis (PCA), and Fisher’s Linear Discriminant to identify typical power load profiles for the author’s institution campus buildings. These methods of machine learning are applied to data collected from the campus and focuses on identifying trends in power usage as well as identify optimal times for charging and discharging of an energy storage system (ESS). Application of the algorithms is carried out using MATLAB to better understand the load profiles of various academic and residential buildings on campus. Bayes Estimation is used to determine optimal times for charging and discharging of an ESS using training sets from the power consumption data. Results from the study show Bayes Estimation yields a high accuracy in state estimation for various sample sizes given a limited amount of training data. Principal Component Analysis is used to determine key features from the data that effectively differentiate between the academic and residential buildings being observed. Key features that are observed through PCA include timescales such as hours of the day, days of the week, and months of the year, as well as power demand readings from each of the buildings’ respective electrical meters. Fisher’s Linear Discriminant is applied to the dataset for a similar purpose to Bayes Estimation, however the algorithm is used to determine peak vs non-peak recordings from the hourly power consumption data. Results from Fisher’s Linear Discriminant method proved to be unsuccessful in discriminating between classes of data. Analysis of the results will be used to further understand where and when ESS can be most effective to reduce peak energy demand from the campus on the local utility grid network. The paper presents the process of applying methods of machine learning to the data as well as the results from the mentioned methods.