Making a workspace comfortable and employees productive isn’t easy. We here at Roby have all experienced this first hand and decided to take on the challenge. To tackle this task, we built an HVAC add-on for Roby, that helps resolve comfort issues by offering a more flexible and realistic way to level thermal comfort in an office.
In this post, we cover our approach and lessons learned in implementing a machine learning system for personalizing thermal comfort in an office space.
In our first iteration of the HVAC add-on for Roby, all users can report hot and cold complaints on demand, either by directly messaging Roby on Slack or by talking to Roby through the Alexa service. Using a dynamical system model, the agent changes the set-points for air flow, chillers, and other HVAC settings. Through these interactions with employees, Roby can predict the required changes per zone and thus reduce complaints, keeping employees focused.
To optimize thermal comfort, we use a neural network that takes as inputs a wide range of features, in order to predict how likely a specific employee will report a hot or cold complaint. These features include:
- Environment status through sensors (e.g., space temperature, humidity, light)
- Environment state of occupancy (e.g., through sensors, or predicted by CO2 and light)
- Tenant behavioral data
- Zone characteristics (e.g., width, length, floor level)
As a positive side effect, we also discovered that our model allows us to find unoccupied rooms and regulate the temperature in those places as well. This way, we can optimize for energy savings in those vacant spots.
As a safety net, if one of the models predicts an extreme change due to the wide range of sensors and task variance, we have set up safe fallbacks. This safe-fallback is especially useful as we expand our datasets across buildings with similar tenant behavior, geographical state and build new models.
We change the environment only if confidence exceeds the set threshold and the employee has given us permission. Today, Roby resolves about 1 in 5 of the complaints and increases the thermal comfort level by cutting the time required to change from hours to seconds.
So what’s next?
First, we’re iterating on the thermal comfort model. The example above considers optimizing for a positive feedback from tenants after the status change, either reporting through a survey the change was helpful or immediately communicating through one of the conversational AI interfaces we offer.
For further predictions such as energy saving and occupancy prediction, however, the baseline-models can and should be further explored. For example, if we observe the data collected jointly, most sensors can be faulty or face the concept drift. In machine learning, the concept drift means that the statistical attributes of the predicted variable, change over time in unforeseen ways. This is a known issue and will likely make the predictions lower in confidence as time passes.
Second, as the safe fallbacks are being set up, we are brainstorming and launching new kinds of machine learning based reasoning. Since the models automatically choose whom to contact after complaints arise from employees, we can explore new ways to engage fewer people, be confident that those parties that are not required will be efficiently held back and bypassed.
If you're interested to learn more, email us at email@example.com
Want to work for our engineering team? Email us at firstname.lastname@example.org