ASHRAE Journal - August 2024 - 36

TECHNICAL FEATURE
FIGURE 4 Optimal combinations of chillers and optimal loads for a scenario of 850
tons (3,000 kW) of cooling, LCHWT of 48°F (9°C) and ECWT of 71.5°F (22°C).
All Chillers Predicted Efficiency
1.4
1.2
1.0
0.8
0.6
0.4
100
200
300
Chiller 1
400
500
Load (Tons)
Chiller 2
Chiller 3
Chiller 4
of the chilled water plant. As shown in Figure 5, there is
a clear trade-off between the pump power and chiller
power. Without knowing the part-load performance of
equipment, the optimal point that minimizes overall
power use would not be obvious. This optimum varies
depending on how much the chiller is loaded, water
temperature, etc. Similarly, there is a trade-off between
the cooling towers and chillers' power use that benefits
from this type of predictive optimization.
All these optimal setpoints and combinations of
equipment can be resolved together as part of a
mathematical optimization framework. The simplest
solution is to predict the overall chiller plant power use
for all possible values of setpoints and all combinations
of equipment and find those that bring the lowest plant
power use. For instance, the algorithm may compute
the power at every combination of chiller's cooling load
between 30% and 100% at a 1% interval, and for each
of these cases, every condenser water flow between
70% and 100% at a 1% interval, etc. This approach is
often referred to as brute-force or exhaustive search.
While valid for fewer variables, it quickly becomes
intractable as the number increases, especially for
real-time controls where the optimal controls strategy
needs to be found quickly on embedded hardware
on-site. There are numerous types of mathematical
optimization solvers, and this is an established field of
mathematics.4,5
One popular method that suits this application well
is the interior-point method because it can handle
nonlinearity and constraints. The latter is key in this
application, as a number of bounds relating to the safe
operating ranges of equipment must be respected,
such as chillers' minimum flows. Specific interior
point and other optimization methods have been
36
ASHRAE JOURNAL ashrae.org A U G UST 2024
600
700
800
900 1,000
carefully designed and tested for various applications,6
and robust open-source software is available.7 These
optimization solvers use similar underlying theory
and techniques as the ML process because the latter is
essentially an optimization problem with the objective
of finding the model that minimizes error between the
prediction and actual data.
The ML and optimal control methods can be deployed
for real-time optimization on-site thanks to progress in
computing hardware. For instance, modern computer
boards available on the market can house quad-core
CPUs of upwards of 1 GHz speed and 1 GB of memory
within a small form factor. Nowadays, such computer
boards are used in a variety of home and industrial
automation applications. They are powerful enough to
host some types of ML and optimization applications,
although care may need to be taken to ensure
algorithms are computationally efficient for largerscale
and more complex applications. HVAC control
is still a rather small-scale case, however, compared
to other ML applications that may rely on millions
of variables, such as image or speech recognition.
As a result, such computers chips are suitable to be
installed in embedded controllers and can be used
to host real-time chiller plant optimization software
directly on premise.
Site Deployment Results
The ML-driven approach presented in this article
has been deployed at dozens of sites around the world.
Measurement and verification studies were conducted
at a number of buildings in Australia, Singapore and
FIGURE 5 Optimal CW flow trade-off between pump power and chiller power.
Conditions: Load 50.0% - LCHWT 44.5°F - ECWT 84°F
140
120
100
80
60
40
20
600
800
1,000
Condenser Flow (gpm)
1,200
Pump Power
Chiller Power
Total Power
Optimum
1,400
Efficiency (kW/ton)
Power (kW)
http://www.ashrae.org
ASHRAE Journal - August 2024

Table of Contents for the Digital Edition of ASHRAE Journal - August 2024
