Abstract

Worldwide, the building sector consumes a significant amount of energy in different stages such
as construction and operation. Depending on the type of energy source used, buildings have a
considerable impact on air pollution and greenhouse gas emissions. To reduce the amount of emissions
from the building sector and manage energy consumption, many tools and incentives are used
around the world. One of the most recent and successful approaches in this regard is the application
of machine learning techniques in building engineering. The increasing availability of real-time
data measured by sensors and building automation systems enable the owner and energy planner
to analyze the collected information and explore the hidden useful knowledge and use it to answer
specific questions such as which parts need retrofit, how much energy can be saved and what would
be the cost. At the building level, machine learning has different applications, such as pattern extraction
and load prediction. Amongst those, load analysis and energy demand prediction are of
specific importance for the building energy managers, as it can lead to a more efficient operation
schedule of energy systems in the building. The analysis of load profiles can give a good overview
of the energy use and user behavior in the building. Detailed load analysis and understanding is an
essential step before the predictive analysis.

In this study, electrical load data from three transformers installed in EV building, Concordia
University and weather data collected from the weather station installed in EV building were used
for load analysis and load prediction. EV building includes two main parts, which are Engineering (ENCS) and visual arts departments (VA). The three transformers considered in this study measure
heating, ventilation, and air conditioning (HVAC) load from a mechanical room (located in 17th
floor of the EV building) in addition to the plug and miscellaneous loads from ENCS and VA departments.
In the load analysis part, the representative daily loads of these three transformers of
the building are studied. The magnitude and trend of daily loads are extracted and discussed. The
average load from 17th floor’s transformer is found to be 1,441 kW during office hours of weekdays
in summer, whereas this load during office time in winter is 991 kW. Note that, this load does not
include the gas consumption, used for meeting the heating load during the winter. Regarding the
plug load from ENCS and VA department, the average load during office hours of weekdays in
summer is 512 kW, and 453 kW, respectively. Moreover, the load reduction during the COVID19
pandemic is studied by comparing the two months (April and May) of 2019 and 2020 for all three
transformers. There was a significant reduction of 42 % for the load of 17th floor between April
2019 and April 2020 (weekdays), while 24% and 40% load reduction was observed for ENCS and
VA transformers, respectively. Based on the results during COVID 19 period, we see that the existence
of people in the building affects the load, but a great part of the load is related to the schedule
and policy of the building. That is why there is a good potential to save energy just by changing the
schedule and plans that systems are running based on.

The second part of the work deals with load prediction using regression analysis and long shortterm
memory (LSTM) model. The importance of input variables for load prediction is evaluated
in the regression section. In linear regression, twenty scenarios are considered. Each scenario is
a different combination of input features. It was found from the results that the best scenario is
when all calendar and weather data are considered as input attributes. The best scenario in winter
has R2=0.29 and MAPE=24.46, while in summer, R2=0.64 and MAPE=10.47. The results are
confirmed with correlation analysis. For this case study, adding meteorological data did not improve
prediction in winter significantly because in winter, gas is used for heating and the considered
data does not reflect it, but in summer, weather variables were of great importance. Also, specific
and unusual events in consumption could be detected with polynomial regression. Regarding load
forecasting, LSTM is used as a deep learning model, which considers the sequential load data and predicts future load for different time horizons. Regarding the size of the dataset and LSTM parameters,
the best performance was obtained for one-year ahead forecasting with R2= 0.75, and MAPE=
10.97. Another result was that the type of load influences the performance of the LSTM model.
Considering different load types, the plug and lighting loads from the ENCS and VA departments
could be better predicted than the 17th floor HVAC load, since HVAC load is affected by weather
variables that are fluctuating and not easy to predict, but plug loads are more related to the schedule
of building. The other influencing factor on prediction performance is the choice of train-set and
test-set. The lowest R-squared belongs to the model that has the year 2019 as test-set. The results
of this project could be useful for building facility managers to adapt and optimize the schedule of
the energy systems and give recommendations to the users to improve energy efficiency.