Optimized Control Using Machine Learning (It’s About Time) The good news is that if or when we develop this technology, we can once again eliminate the arduous task of line by line programming and squeeze every drop of energy waste out of buildings.

John Pitcher CEO Weber Sensors john.pitcher@captor.com

Articles

Interviews

Releases

New Products

Reviews

Editorial

Events

Sponsors

Site Search

Newsletters

Archives

Past Issues

Home

Editors

eDucation

Training

Links

Software

Subscribe

Unlike some of my contemporaries I was never a fan of pneumatic controls and have no remorse for their passing. Sure I had a lot of contracts with customers to periodically calibrate their systems and made good money doing it, but I always knew full well that it was highly unlikely those systems would stay calibrated till the next maintenance interval. In fact, I would wager to say that some would not stay calibrated for more than a few hours.

When I was first introduced to the Andover Controls Sunkeeper back in the late 1970s, I was hooked. The Sunkeeper was the first truly DDC system I had seen that could replace pneumatics. Just add, what was then a blazing fast 300 baud modem, and you could remote into the building through a dumb terminal. Want to add a couple of days’ battery backup? Buy a Sears Die Hard car battery (I kid you not). The “drum” programming was a bit awkward to learn but none the less if you were smart and creative enough it could do things that no pneumatic system could match. More importantly, it was far more accurate and reliable than pneumatics and practically bullet proof.

I was convinced that the Andover system was the end all be all when in the early 1980s I was introduced to the Novar Controls system. Many may not be aware, but Novar was the first building automation system that I know of to incorporate self-learning control algorithms. The control algorithms eliminated the need for PID loops and the manual tuning that goes along with them. Not only were Novar’s control loops self-learning their optimized start-stop program was as well. What took me days to program in the Andover System took hours on the Novar and best of all, there was no de-bugging needed. Because of its self-learning algorithms, the Novar system was programmed via simple menu driven commands. No programming language to learn, easy to use. Novar proved that the more intelligence you can put into the control method the easier it is to program and deploy.

There was, however, one fallibility with the Novar system that is worth mentioning. The self-learning algorithms were great as long as the systems it was controlling were working correctly. On the other hand, if there was a fault, all hell broke loose. For example, if a chiller was running inefficiently and the cooling process was slower than normal, you might find the optimized start program starting the building up at mid-night. The conclusion being, if systems have a fault, advanced algorithms can make matters worse and waste more energy than systems without intelligence.

It has been more than three decades since Novar introduced self-learning algorithms for control loops in buildings and where are we today?

Building automation systems are still pretty complicated to program and deploy, and we continue to treat the control of buildings as a series of PID loops instead of what they are which is a unified, holistic system.

In the 1980s we started designing mostly variable speed buildings. One would think that by now automation systems would be able to automatically determine if it makes more sense to increase fan speed on the air handlers or reduce chilled water temperature, just as a simple example. Controlling systems holistically to get to the lowest energy usage is surprisingly missing.

It isn’t that we lack the knowledge, these types of machine learning algorithms have been around for decades and are always being improved and optimized. In the 1980s the microprocessors used in the early building automation systems had a fraction of the power of those today but artificial intelligence (AI) algorithms such as Simulated Annealing which probably could have solved this problem were. Now I don’t know if the microprocessors back then would have been powerful enough to run those algorithms but certainly the microprocessors we have can today. Furthermore, newer better-optimized machine learning algorithms such as Extended Compact Genetic Algorithms have gained favor over simulated annealing, and there is little doubt that modern microprocessors have the horsepower needed to run them…..but hold the fort.

As Novar discovered in the 1980s, if the machines you are trying to optimize are not performing correctly, machine learning can easily make things much worse. It is much like the adage that an accountant can make a mistake of several dollars whereas a computer can make a mistake of millions. This is a challenge to overcome.

It seems that deploying an effective fault detection program is clearly needed before implementing machine learning. The fault detection program would then be tasked not only to make maintenance more efficient but also to qualify whether the machine learning algorithms would and should be allowed to execute. Without this precaution, I believe machine learning as it pertains to optimizing building control has the potential to cause more harm than good.

The good news is that if or when we develop this technology, we can once again eliminate the arduous task of line by line programming and squeeze every drop of energy waste out of buildings.     
 

About the Author

John Pitcher is the CEO of Weber Sensors working on creating the next generation of sensors for the HVAC industry. Mr. Pitcher was the founder of Scientific Conservation one of the first cloud based fault detection companies and has held many executive positions in his 45 plus years in the HVAC industry.  He can be reached at john.pitcher@captor.com