ASHRAE Journal - August 2024 - 33

TECHNICAL FEATURE
results in lower pump power use but is also expected to
increase the chiller power use, as discharge pressure on
the chiller increases with a lower condenser fl ow (when
all other independent variables stay constant). Similarly,
decreasing the leaving cooling towers' temperature
setpoint reduces chiller power use but increases cooling
towers' power use. It is not always obvious which
setpoint value will result in the greater overall energy
savings in the plant, and it may vary depending on
conditions from site to site and over longer periods of
time as the performance of equipment may degrade.
Another area of interest is fi nding the optimal number
of chillers to run at any given time. A dynamic and
predictive approach would be benefi cial, as this may
vary based on cooling demand, water temperatures but
also the part-load performance of each chiller and which
chiller is available or offl ine due to fault or maintenance.
Another potential obstacle for deploying ML systems
in chiller plant controls is that they are often designed
to operate in the cloud, and it can be challenging to
interoperate with the on-site BAS, while increasing
the risk of downtime due to connectivity issues. On
the other hand, controllers deployed on-site to control
the plant have limited computing and software logic
capabilities and would likely not be suitable to host ML
software programs. With recent advances in embedded
computing hardware and numerical methods, it is more
viable today to carry out some level of ML computations
directly on-site, without requiring a cloud platform.
Machine Learning Method: An Overview
To fi nd the optimal setpoints and combinations of
equipment, the fi rst step of the proposed approach is to
apply supervised learning to learn the performance of
the equipment from sensors- and meters-recorded realtime
data. Supervised learning covers a broad range of
techniques used to train a model on a data set to attempt
predicting outcomes. In this case it can be applied to
learn the power use or coeffi cient of performance (COP)
(in kW/kW or kW/ton) of chillers, pumps and cooling
towers for various conditions of cooling load, water fl ows
or water temperatures. Once trained with recorded
site data, the model is able to predict the power use
of the chiller plant in various situations and capture
the interaction between each piece of equipment. The
process is repeated regularly, for instance once a day,
to account for new data and continuously learn the
FIGURE 1 Chiller predicted power (surface) compared to actual power (dots).
100
90
80
70
performance profi le, including degradation over time,
that may affect the optimal sequence.
An optimal control process can then leverage ML
performance models and fi nd the best control strategy
dynamically and with different outcomes depending
on conditions and learned equipment part-load
performance. The overall process is autonomous and
can operate in real-time without human intervention.
It learns the best sequence of operation, continuously
updating based on equipment performance and site
conditions.
Machine Learning in Action
The supervised learning approach is able to capture
the power use of a chiller as a result of the cooling load,
the leaving chilled water temperature (LCHWT) and
the entering condenser water temperature (ECWT), as
displayed in Figure 1, with the predicted power compared
to the actual recorded data points from a site. In Figure 1,
the entering condenser water temperature and leaving
chilled water temperature are combined in one for ease
of visualization on a 3D graph.
It can also capture more complex relationships, like
the impact of condenser water (CW) fl ow in a case with
variable speed condenser pump. Figure 2 displays the
effi ciency (kW/ton) profi le of a water-cooled chiller
that was learned from the system observing the chiller
operate on-site through real-time data gathering.
At a lower condenser water fl ow, chiller effi ciency
noticeably decreases.
The effi ciency (kW/ton) curves of chillers are fairly well
A U G U ST 2 0 2 4 ashrae.org ASHRAE JOURNAL
33
Load (Ton)
10 15 20 25 30 35 40
ECWT-LCHWT (°F Difference)
Power (kW)
50 100 150 200 250
http://www.ashrae.org
ASHRAE Journal - August 2024

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