The Data Science Behind Room Temperature Optimization – BuildingIQ

Blog

Partner Spotlight – Aegis Systems
April 15, 2019
Smart Grid Edge Analytics Workshop: The Role of Building Analytics in Utility Strategy
June 21, 2019
Show all

The Data Science Behind Room Temperature Optimization

Room temperature optimization refers to the planning of a future room temperature setpoint profile that maintains comfort and reduces energy, thus lowering energy costs and peak demand. In a Heating, Ventilation and Air Conditioning (HVAC) system, energy is consumed to maintain, among other things, a comfortable building room temperature. The optimization assumes that comfort is subjective, there is no single temperature that is best for all building occupants in all conditions and there is a comfortable temperature range where the percentage of satisfaction does not vary with any statistical significance.

Recently, we published a case study “Operational & Energy Optimization at University in Australia­” about an ongoing project we’re working on at a reputable university in Sydney. In this article, I’m presenting a deeper view of this case study into the performance of our Predictive Energy Optimization (PEO) at one building at the university.

The building is multipurpose, includes labs, offices and common areas. The building’s HVAC system consists of many Fan Coil Units (FCU), which consume chilled and hot water to modulate the supply air temperature and hence maintain the room temperature of a localized building zone. The chilled and hot water is supplied from a central chiller plant and a central boiler plant that also supply other buildings. In September 2018, PEO was gradually turned on and it has been running for about 50% of all FCUs during office hours.

The method for this deep dive was to compare the energy consumption and the operation of the HVAC system using PEO vs. without PEO; we refer to the former as the “reporting” period and the latter as the “modelling” period.

In this article, I’m presenting one set of results from the case study. In this set, we measured the average chilled water valve (CHWV) position of all sixth level FCUs that are under PEO optimization during office hours. We compared the average CHWV of February 2019 (reporting period) to that of March, April and May 2018 (modelling period). Ideally, we would have liked to compare it with January, February and March 2018 but the BMS data we have starts from March 2018. Due to colder weather in the modelling period compared to the reporting period, a model fitted to the data in the modelling period tends to underpredict the average CHWV in the reporting period. This bias means that we are more confident in any saving that we might calculate. With the maximum comfort temperature set to 24°C (75.2°F), which is 2°C (3.6°F) above the fixed setpoint of 22°C (71.6°F) in the modelling period, the average chilled water valve position is reduced by 44.8%. See Figures 1 to 3 for a detailed comparison between the reporting and the modelling period.

Figure 1. This figure shows the saving measured on the average CHWV of the sixth level FCUs that are under PEO optimization during office hours. The first subfigure shows the accuracy of the model and the second subfigure shows the difference in the actual power and the expected power that is based on the data in the modelling period. The average CHWV is reduced by 44.8%. However, the reduction at night is not due to PEO optimization.

Figure 2. In this figure, we took one day in the reporting period (the dotted black line), looked for similar weather days in the modelling period (solid non-black lines) and made comparisons on those days. The three subfigures on the left compare similar weather days: the first subfigure shows how similar the outside air temperature profiles are, the second subfigure compares the average CHWV and the third subfigure compares the average zone temperature. We can see that, in the reporting period, there was a change in building operation during the night: there was less chilled water consumption and the zone temperature was allowed to drift much higher. This is not due to PEO but it is why there was more chilled water consumption in the morning to bring the zone temperature down to within the comfort range. In this reporting period, the comfort max was set to around 24°C (75.2°F). If we were to bring the zone temperature down to 22°C (71.6°F) as in the modelling period then the FCUs of this level would have consumed even more chilled water. The three subfigures on the right are just to show days on which the weather was not similar to that one day in the reporting period and how the results contrast the subfigures on the left.

 

Figure 3. See the description for Figure 2. In this figure, we picked a different day in the reporting period. These figures clearly show the effect of zone temperature profile on the average CHWV, which is really a proxy for the chilled water consumption. On this day, the optimized room temperature stayed close to the comfort maximum of 24°C (75.2°F). One argument against room temperature optimization is such that if the savings are mostly achieved by raising or lowering the room temperature to a different fixed setpoint level then one can always manually achieve this without the complex mathematical optimization achieved through PEO. However, given the outside air temperature, the decision to have the zone temperature kept at or drift toward the comfort maximum or minimum is not always straightforward. The modelling process in our PEO algorithm learns and estimates the thermal load of the building in order to make this decision.

This deeper view on this particular case study shows the impact of Predictive Energy Optimization (PEO) both on the whole building meters and on the per-zone HVAC mechanical operation. The impact was positive both in terms of energy saving and comfort. The Measurement and Verification strategy was adapted to the commercial project, which presented unique challenges other than the ones normally seen in a more controlled experiment setting.

Based on this deep dive into this project, we believe the following factors contributed to its success:

  1. A wide and adjustable comfort range was agreed upon by working closely with the University’s onsite team,
  2. An efficient HVAC system that allowed PEO to have good control of room temperatures, and
  3. Reliable power meter data that allowed us to constantly evaluate the performance of the HVAC system especially in relation to PEO.

 



Dr. Rui (Ray) Xu is Data Scientist at BuildingIQ. During his career, he has encountered many challenges in the transfer of human knowledge into machines, the interpretation of results of algorithms into human intuition, and the verification of evolving strategies. This experience has helped him to solve some of the most difficult and interesting problems in the industry. He describes himself as a bit quiet but also very energetic and assertive when there is a mathematical, simulated and/or experimental proof behind the scenes.

 


Adam Benson is VP of Engineering and DevOps at BuildingIQ. Adam provides the guidance and leadership to bring to market a platform of services that enable today’s buildings to be smarter and greener.