Paper Conference

Proceedings of eSim 2022: 12th Conference of IBPSA-Canada


Evaluation of Data-driven Models for Short-term Electrical Load Forecasting in Office Buildings

Farid Bahiraei, Hadia Awad, Araz Ashouri
National Research Council, Canada

Abstract: Load forecasting algorithms play a key role in the successful management of building operation schemes such as peak management and time-of-use optimization. Such implementations require reliable historical energy use data to train the models, which might be unavailable for newer buildings, or in cases of major interruptions in data collection. In this paper, we present data-driven approaches to predict electrical loads for commercial buildings with limited training data. The models are based on temporal, autoregressive, and exogenous variables and are designed to reduce the impacts of the COVID-19 pandemic on their prediction performance. Model validation is performed with data from an existing commercial building connected to a smart metering system. The results indicate that, compared to simple and linear models, non-linear models provide enhanced and acceptable forecast accuracy even when trained over relatively short periods of time, and despite changes in the energy use pattern during the pandemic.
Keywords: electrical load forecasting, smart metering, machine learning, commercial buildings