1. Personalized HVAC Control System
Mark Feldmeier and Joseph A. Paradiso
Presented By: Debarun Das

2. Outline Motivation System Overview Algorithms Experimental Setup
Evaluation
Conclusion and Shortcomings
Further Readings

3. Motivation Lack of user comprehension of the behavior of HVAC systems
Largest consumer of energy in buildings are the HVAC systems
26.1% and 53.4% of total energy usage in residential and commercial applications, respectively
Usage of new, more efficient ventilation technologies are expensive
Need to use existing technology more efficiently
Lack of user comprehension of the behavior of HVAC systems
Users tend to consistently over-turn thermostat dials in uncomfortable conditions
Leads to energy wastage

4. System Overview Portable Nodes Control Nodes
Sense parameters relevant to human temperature comfort
Temperature, Humidity, Light Level, Activity Level
Allows user to input current comfort state – hot, cold, neutral
Fixed versions gather internal and external local climate data
Control Nodes
Actuates air sources (windows and air conditioning dampers)
Have a motor that opens/closes the corresponding mechanical component

5. Room Node Central Network Hub
Network Coordinator – Receives data from portable and control nodes
Helps infer locations of portable nodes from the RSS value
Also sends local conditions (temperature, humidity etc) to the central network hub
Central Network Hub
Computer that receives all data over ethernet
Processes data according to control and comfort algorithms
Issues motor control commands to room nodes

7. Control Algorithms Control Module Location Module Thermostat Module
Receives setpoints from other modules
Location Module
Locates Portable Nodes
Based on strongest RSSI from room nodes
Majority vote of past 3 readings smooth out the data
Thermostat Module
Normal – Room occupied but portable node not worn
Setback – Room unoccupied. Temperature raised
Active – Room occupied by people wearing portable node
Portable Module
Keep tracks if a user is active
Uses temperature
Uses windowed mean and variance of piezoelectric motion sensor
Collects user comfort level data on button press
Room Module
Polls Location and Portable Modules
Gives control to thermostat module when needed
Detects room occupancy by using PIR motion sensor

9. Comfort Algorithm
Need to measure Comfort Distance – how far he/she is from being comfortable
Constraints
Insufficient data labels
Reduced Data Set
Indeterminate Data: Lack of even distribution of hot and cold labeled data
Solution
Robust Algorithm
Does Not Overfit

10. Fisher’s Linear Discriminant
Maximizes the separability among known categories
Choose a decision boundary that maximizes
𝑱 𝑭𝒊𝒔𝒉𝒆𝒓 = 𝑺𝒆𝒑𝒂𝒓𝒂𝒕𝒊𝒐𝒏 𝒐𝒇 𝒑𝒓𝒐𝒋𝒆𝒄𝒕𝒆𝒅 𝒎𝒆𝒂𝒏𝒔 𝑺𝒖𝒎 𝒐𝒇 𝒘𝒊𝒕𝒉𝒊𝒏 𝒄𝒍𝒂𝒔𝒔 𝒗𝒂𝒓𝒊𝒂𝒏𝒄𝒆𝒔
Representative training points have extreme values
Normalization of Comfort Distance of a user
Comfort Distance of the user divided by the mean Comfort Distance of all users

11. Experimental Setup Performed at MIT Media Laboratory
Ten people, mostly graduate student population
Four offices and common room area equipped with control nodes and room nodes
Study conducted in three phases.
Phase 1
No actuation
Repairs done
Hardware Evaluation
Phase 2
Questionnaire
Portable Node Feedback
Phase 3
Control System tested
Periodic Surveys
Week 2, 3, 4

12. Evaluation - Energy Metrics
Energy usage measurement by air flow sensors in VAV boxes
Average of Week 2 and Week 3 shows 75% decrease than Phase 2
Actual savings are much smaller
Room 36 and 40 also cooled by 8 other non- controlled VAV boxes
HVAC did not run properly in Phase 2 for all rooms
Room 44 analyzed in detail
Energy saving of 11.8%
Average energy saving 24%

13. Evaluation - Comfort Metrics

14. Conclusion and Shortcomings
Managed to achieve energy savings of 24% over standard HVAC systems
Shows considerable improvement in overall user comfort level
The outside air is used for cooling only as the experiment is performed in the summer months
The activity level thresholds are picked manually for different users after six weeks of user data
Accelerometer could have provided more accurate activity level data
Survey not fully reliable
Small number of users
Users mostly graduate students
Should be performed across people of different age groups

15. Further Readings
MarcoPritoni, KiernanSalmon, AngelaSanguinetti, JoshuaMorejohn, MarkModera, “Occupant thermal feedback for improved efficiency in university buildings”
Vitor Mansur, Paulo Carreira, Artur Arsenio, “A Learning Approach for Energy Efficiency Optimization by Occupancy Detection”
Alimohammad Rabbani, S. Keshav, “The SPOT* Personal Thermal Comfort System”