Find Us On:

Linked In


Join Our Mailing List


We won't spam you or ever share your information.

Researchers optimize air conditioning units for max efficiency


In Singapore, more than 40 percent of a building’s total energy consumption can be attributed to its chiller plant, according to ADSC Senior Research Scientist Zhenjie Zhang. By using new technology developed as a collaboration between ADSC and a Singapore air-conditioning company, Kaer, companies could save thousands of dollars per month on their energy bills.

A Kaer chiller plant
A Kaer chiller plant
Poor efficiency in an HVAC system is often due to the excessive overhead and technical difficulty of manually tuning a chiller plant, which is the part of an air conditioning system that is responsible for keeping the air cool. ADSC and Illinois researchers are working with Kaer, who owns and operates numerous chiller plants across the island, to develop an advanced optimization and data management platform that can be easily used in any building without expert knowledge. By using the researchers’ machine learning algorithms, some plants have already realized savings of five to 20 percent.

The collaborative project brings together experts in the industry in computer science, including ADSC’s Zhang and University of Illinois Computer Science Professor Emeritus Marianne Winslett, Nanyang Technology University Computer Science Assistant Professor, and ADSC affiliate Xiaokui Xiao, and in mechanical engineering, including University of Illinois Mechanical Science and Engineering Professor Andrew Alleyne and Illinois graduate student Bryan Keating. The project is funded by a two-year, $1.2 million grant through Singapore’s Building Construction Authority, as part of the Singapore Green Buildings Innovation Cluster (GBIC).

“One of the things we’re excited about is that at Kaer we design and operate the chiller plants, so we know the problems and how expensive these systems are,” said Chai Kok Soon, director of Research and Development at Kaer. “By bringing together all these people with different backgrounds, we can hopefully use the data we’ve collected to help solve the problems.”

The goal is to use the chiller plant in a way that uses the least amount of energy possible while keeping the building at the proper temperature. Developing a solution is difficult because of the complexity of the plant system and is compounded by external factors such as varying weather patterns, building usage, and the fact that different buildings may have completely different characteristics, which doesn’t allow for easy transfer of technology.

Machine learning provides the theory to predict independent variables, such as power and cooling load, from dependent variables for chiller plants. However, the major challenge is to apply the theory to implement a real-time, autonomous learning system. The learning-based energy management system needs to perform automatic optimization and baselining to evaluate energy saving. Kaer developed a software platform to implement the theory with data visualization, optimization, real-time control and Internet of Things modules to do continuous training, testing, evaluation and learning of actual chiller plants.

By partnering with Kaer, the researchers can test their optimization algorithms on a live system, as well as use years’ worth of data that Kaer has collected from its plants -- which are in places like shopping malls, schools, and businesses -- to develop their methods.

“Working with ADSC and Kaer has been very valuable to us in obtaining real world data on complex systems operating in challenging practical environments,” Alleyne said. “It helped us refine our modeling and simulation tools. It also gave us good insight into the effectiveness of our algorithms and ways we could modify them to meet the real-world situations.”

According to Zhang, the chiller plant industry currently requires experts to manually alter chiller plants for optimal performance and most companies don’t have full-time engineering support.

“Unfortunately, even though these engineers can control the plants in an efficient way to save more energy, it’s very difficult to train these engineers,” Zhang said. “That’s why machine learning can be a good addition to the current technology.”

Zhang added that this is also a great time for this research because sensors have recently become an essential part of chiller plants, which allows researchers to collect and analyze the plant’s data, allowing them to understand the plant’s behaviors.

The researchers have found that the platform consistently outpeforms expert engineers when it comes to cost savings.

“Since September, we have achieved savings of four to nine percent or even more, compared to some of our most well-run chiller plants,” Kok Soon said.

Kok Soon added that they have seen five to 20 percent improvement on many chiller plants, depending on the expertise of the energy experts, control strategy, equipment conditions, and frequency of adjustments. This leads to round-the-clock savings for the companies, as well as freeing up engineers for more high-value work, such as diagnostic problem solving.

The researchers are working on a platform that will enable their work to use on a large scale, as well as identify, diagnose, and send alerts when something goes wrong with a sensor.

“We have a few plants that are using this technology now and we are hoping to make this technology available to the public soon, so many companies will be able to do optimization without human expertise,” Kok Soon said.