A Framework to Predict Variability Characteristics in Building Load Profiles

Authors

SAM MOAYEDI, University of Nebraska–Lincoln
ANDREW PARKER, National Renewable Energy Laboratory
KEVIN JAMES, University of Nebraska–LincolnFollow
XIAOYUE CHENG, University of Nebraska OmahaFollow
MICHAEL HEMPEL, University of Nebraska–LincolnFollow
HAMID SHARIF, University of Nebraska–LincolnFollow
MAHMOUD A. ALAHMAD, University of Nebraska–Lincoln

Date of this Version

1-27-2023

Citation

S. Moayedi et al.: Framework to Predict Variability Characteristics in Building Load Profiles. VOLUME 11, 2023. DOI: 10.1109/ACCESS.2023.3239328

Comments

Used by permission.

Abstract

The expansion of Advanced Metering Infrastructure (AMI) has provided building operators and researchers detailed information on building energy consumption. The majority of AMI systems, however, record data at relatively low resolutions of 15, 30, or 60 minutes, due to cost, storage and bandwidth limitations. Emerging applications in power flow analysis, Quasi-Static Time-Series Simulation (QSTS), smart grid integration and load matching, however, require data at higher resolutions. Short-term energy demand can deviate significantly from long-term averages, with an unknown magnitude and frequency when only low-resolution load profile data is available. This paper presents a novel data-driven approach to predict characteristics of the missing high-resolution information in a low-resolution signal, applicable to both measured and modeled building load profile data, utilizing machine learning regression algorithms. In the proposed framework, the relationship between characteristics of high-resolution and low-resolution signals is learned from the decomposition and characterization of a subset of high-resolution building data. This paper validates the underlying hypotheses and methodology of this approach through a single-building case study, training a variety of machine learning models on one year of data, and using the resulting models to predict high-resolution characteristics in a different year. An Ensemble Tree regression model demonstrates a high predictive accuracy (R² of 0.79-0.92) for several statistical metrics of the high-resolution load profile. These results support the broader potential for leveraging low-resolution information to accurately constrain predictions of missing high-resolution information in building load profiles, which may greatly increase the utility of both measured and modeled data in many practical and research applications. Generalizing such models will require analysis of high-resolution data from a diverse set of building types.