# modeling for HVAC system

May 21, 2015
Algorithms

Report Abuse
This experiment uses regression analysis to predict the power consumption of HVAC system in unit time.
Smart Energy technology plays a key role in efficient building operation and also has a close relationship with emissions reduction. We try to tackle this technology from one angle, which is the efficiency of HVAC (heating, ventilating and air conditioning) system inside a building. As an example, heating and cooling accounts for half of the energy use in a typical U.S. home, making it the largest energy expense for most buildings. To improve efficiency of HVAC system, we started by building a regression ML model which could predict the power consumption of HVAC system in unit time, based on several input features. To be more accurate, the problem we are trying to solve is - how does the HVAC power consumption in unit time depend on values of several other factors found during the same period. Usually it takes quite some time and effort to learn a deployed HVAC system, even for domain experts and HVAC operators, because of different building types and environments. However our introduction of machine learning obviously accelerated this learning process. The target building is equipped with a water-cooled HVAC system. It is composed of several subsystems like cooling tower, chiller, condenser. Basically there are two water flow loops in a water-cooled HVAC system: ![Water flow loops in HVAC system][1] **Chilled water loop** (“Very cool water” in the diagram above) runs through the building, picking up heat along the way, which is circulated through the chiller to be cooled. **Condenser loop** (“Cool water” in the diagram above) runs through a cooling tower, which uses evaporative cooling to cool this piped water, and then circulated back to the chiller to take the heat from chilled water loop. We first chose quite a few indexes like water flow rate, water temperature, water pressure of different locations based on our discussion with HVAC operator and our understanding. We then used feature selection to reduce dimensions. Finally four indexes are selected as input features for our regression model to predict power consumption: - **chT1**: Temperature of water in chilled water loop before going through Air-handling unit. - **chT2**: Temperature of water in chilled water loop after going through Air-handling unit. - **chFlow**: Flow rate of water in chilled water loop. - **coT1**: Temperature of water in condenser loop before going through chiller. In our project, we have created this regression model using AML and then published a prediction web service API for this model. After that, we used the API to visualize prediction result under certain value combinations/ranges. ![Visualization of HVAC regression model][2] In the graph above, chT1 is set to a range between 8°C and 14°C, chFlow is set to a range between 1000 LPM and 4600 LPM, other features are set to fixed value. A rapid power consumption rise could be observed from the graph when chFlow increases from below 3600 LPM to above 3600 LPM. This phenomenon is usually described as capacity failure in HVAC system. It means chiller has almost reached its limit for cooling capacity, and a blind increase of water flow will only waste more power and no more heat will be picked up. It is worth mentioning that we did find several occurrences of capacity failure when chFlow>3600 after re-examining real historical data. Based on these observations, we define 3600 LPM as a threshold in this HVAC system. Whenever chFlow goes above 3600, an alert will be triggered in the control system of this building, and a backup chiller will be activated to alleviate this capacity failure situation. It might be confusing that why prediction is needed instead of investigate the real HVAC system directly, because we could get the power consumption in unit time directly with certain value combinations of observed chT1/chT2/chFlow/coT1. The problem is we might need millions of tests on real building to explore all combinations of these variables. Instead, by adopting our trained regression model, we could predict power consumption under any untested combinations and discover the most efficient “patterns” as explained above. You could choose other features of HVAC system which might be more suitable for a different setup. Note that we didn't use real data in this dataset due to non-disclosure agreement. Sample code for consuming AML batch API and drawing the 3D graph above is shared on GitHub: https://github.com/shuxianc/hvac [1]: http://iotcatml.blob.core.windows.net/misc/icon.png [2]: http://iotcatml.blob.core.windows.net/misc/3D.png