Page 40 - Climate Control News magazine Oct-Nov 2022

Review
Controls strategy to reduce energy use
“The proposed approach is to learn the part- load performance of Chillers and Condenser Water pumps from recorded data, with a Machine Learning solution that is intended to run autonomously on site, without human interaction.
“These models are then used to generate a Digital Twin of the plant and applied in a predictive optimi- zation framework with the objective of minimizing the plant power usage by varying the Cooling Load per chiller, the Condenser Water (CW) flow set- points, and the combinations of chillers to run at any given time and sets of conditions.”
Berger said the modeling approach was vali- dated with three separate anonymised sites, in Australia and overseas, and the Machine Learning models showed high accuracy.
This is illustrated by the good fit between the predicted and actual power usage of chillers and CW pumps.
He said the whole solution, including the pre- dictive optimization algorithm, is intended to operate automatically on site, on the edge, on low-cost embedded controls platform, and do not require a cloud component.
“This allows for lower cost due to the absence of on-going fees, and increased relia- bility and cybersecurity. An Interior-point method for Non-Linear Programming was used and the solution is able to run on a low- cost Raspberry Pi 3+ compute module-based platform with a 4-core CPU and 1GB of RAM,” Berger said.
After prototypes and site trials, the solution was eventually implemented in the chiller plant. Optimization solution PlantPRO was used with a full-suite of GUI and flexible configura-
tion for various chiller plant configurations.
CHILLED WATER PLANTS are widely used globally for air-condition- ing and refrigeration appli- cations, and account for a large proportion of the energy used in commercial buildings.
In a presentation at ARBS, the head of R&D at Conserve It, Michael Berger, presented Machine Learning techniques to develop an optimal controls strategy to reduce energy use while maintaining the
required chilled water production.
The strategy identifies in real-time the opti-
mal number of chillers, cooling load distribution amongst chillers and condenser water flow set- points that minimize power usage of the plant.
Berger said chilled water plants have become more complex, partly due to the proliferation of Variable Speed Drives (VSD).
“The plant is an interconnected system, and varying the speed of one machine often impacts the energy usage of other machines, inducing trade-offs that can be challenging to predict,” Berger said.
“Our hypothesis was that by using Machine Learning and predictive optimization, it might be possible to find the optimal control setpoints that minimize the energy usage, and that sig- nificant energy savings would be achieved.
LEFT: Michael Berger
ABOVE: Chilled water plants have become more complex to manage.
TABLE - Chiller & Pumps model learning results with site data
Site A
Site B
Chiller n°
1
2
1
2
MAE (kW)
7.1
4.5
6.3
6.4
MAE/Mean (%)
5.17
4.98
5.49
4.32
R-squared
0.962
0.978
0.957
0.961
Sites
Variables
CHW Plant Savings (kWh)
CHW Plant Savings (%)
Site A
CW Flows
15,948
3.6
Site B
CW Flows and Loads
7,358
7.1
Site C
Loads
4,056
5.1
Pump n°
1
2
1
2
MAE (kW)
0.56
0.31
0.38
0.41
MAE/Mean (%)
4.78
3.47
3.57
3.66
R-squared
0.978
0.988
0.958
0.947
Sites
Plant configuration
Climate
Site A
2 multi-centrifugal-compressors chillers rated at 1,700 kW of refrigeration
Subtropical
Site B
2 centrifugal-compressors of 1,800 kW of refrigeration
Tropical
Site C
2 chillers of 3,000 kW & 1 chiller of 1,000 kW of refrigeration
Temperate oceanic
CLIMATECONTROLNEWS.COM.AU
40