ASHRAE Journal - March 2023 - 28

FEATURE
building design. For example, Neto and Fiorelli used a
NN model to predict the thermal performance of the
administration building at Brazil's University of São
Paulo. Neural networks showed agreement between
energy consumption forecasts and actual values, with
an average error of ~10%.9 In building lighting, the U.S.
Department of Energy reported that artificial lighting
use in 2015 was estimated to be 17.1% of the total energy
consumption, equal to 29.1 billion kWh and 183.6 tons of
carbon dioxide in one year.
To reduce energy consumption required to light buildings,
researchers examined using various AI algorithms
to simulate daylighting performance of buildings that
can reduce dependency on artificial lighting, saving
energy and carbon emissions. Several studies have
proven that AI can accurately and effortlessly simulate
light in a building, reducing the time and cost of these
investigations in the early phase of building design.10-12
Although these studies demonstrated that AI models
are a time-efficient alternative to the existing timeconsuming
and computationally expensive simulation
tasks, researchers have not fully integrated AI with
designers' modeling and simulation environments. This
leads to limited access to AI models for architects without
technical and programming experience. Therefore,
finding ways to fully integrate AI algorithms with existing
geometric modeling tools and hide complex algorithms
from non-programmer users is crucial, enabling
access to all professionals regardless of technical skills.13
AI for Building Operations and Maintenance
Besides design, AI can also improve the operation of
buildings. Advanced control systems ensure that buildings
operate optimally to reduce energy use while maintaining
comfort. Model-based control systems such as
model predictive control (MPC) require precise physicsbased
mathematical models to achieve this, which in
practice are difficult and expensive to obtain.14 AI methods
can help alleviate this need for precise models.
For example, reinforcement learning (RL) approaches
can learn the model from the data alone and improve
the operations of the buildings adaptively over time.15
The most sophisticated approaches to date, e.g.,
RL-MPC, combine the best of both worlds: they use
simple physics-based models to ensure reliability and
explainability of the method, while using AI for online
improvement and adaptation.16 Also, physics informed
28
ASHRAE JOURNAL ashrae.o rg M A R CH 2023
(or constrained) neural networks (PINNs) can ensure
that black-box models such as neural networks do not
produce outputs inconsistent with physical laws.17
Outside of optimal control, AI methods can also be used
to monitor and maintain the building operations. For
example, from past data, typical behaviors of the building
can be obtained, e.g., daily load profiles of the whole
building or operations of subsystems, using clustering
methods like k-means. Once established, during operation,
the current load profile can be compared to the typical
load profiles, and deviations from it can be flagged
for the maintenance team to be investigated further.18
The Data Availability Challenge
One big premise of AI-or really any data-driven technique-is
that it can scale an approach (e.g., on operation
or maintenance) to many buildings without incurring
additional manual work. But one big challenge
is being able to demonstrate such an approach. There
is not typically a lot of freely available data that would
allow for the exploration of AI methods. One reason for
this is that data is not always stored, especially in older
BMS systems. Another barrier is that even if data is available,
the building industry does not yet have a data sharing
mindset like those developed in other industries.
Part of this challenge is simply the unknown of why one
should share data, and another concern is that one may
not feel comfortable sharing potentially intrusive data.
Fortunately, there have been positive developments
in this field in recent years. In 2019 there was an energy
prediction competition ( " The ASHRAE Great Energy
Predictor III " led by Clayton Miller) that had the goals
of collecting as much data as possible and demonstrating
the utility of a large dataset to benchmark advanced
methods against each other.19 Another data collection
effort was undertaken by several national labs
and resulted in the Benchmark Datasets of Building
Environmental Conditions and Occupancy Parameters.20
AI Workforce Development
Developing high-performance buildings will be one of
the biggest challenges as the world's city's grow. AI technologies
can provide fine-grained insights into building
performance and operation, air quality, public health,
mobility and more to deal with challenges dense urban
populations face today and in the future.
Creating truly effective and transformative AI
http://ashrae.org
ASHRAE Journal - March 2023

Table of Contents for the Digital Edition of ASHRAE Journal - March 2023
