Peter Xiang Gao, S. Keshav University of Waterloo
HVAC Energy use Buildings use 1/3 of all energy 30-50% of building energy is for HVAC Can save energy by changing temperature setpoint: 1 o C higher when cooling ≈ 10% saving 1 o C lower when heating ≈ 2-3% saving
Focus of this work Consider a single office heating system in winter Assume Thermal isolation Personal thermal control (heater)
Personal Office Thermal Comfort Management Office Corridor
SPOT+: A Smart Personalized Office Thermal Control System Occupancy Prediction Learning-Based Modeling 500W f () + 1 o C -> Personal Thermal Comfort Evaluation Arrive officeLunch Setpoint Scheduling
SPOT+: A Smart Personalized Office Thermal Control System Occupancy Prediction Learning-Based Modeling 500W f () + 1 o C -> Personal Thermal Comfort Evaluation Arrive officeLunch Setpoint Scheduling
Predicted Mean Vote (PMV) model Air Temperature, Background Radiation, Air Velocity, Humidity, Metabolic Rate, Clothing Level ColdCoolSlightly CoolNeutralSlightly WarmWarmHot ASHRAE Scale
SPOT [1] Clothing Sensing Microsoft Kinect: Detects occupancy Detects location of the user 5° infrared sensor: Detects users’ clothing surface temperature _______________________________ [1] P.X. Gao, S. Keshav, SPOT: A Smart Personalized Office Thermal Control System, e-Energy 2013 WeatherDuck: Senses other environmental variables
Clothing level estimation Estimate clothing by measuring emitted infrared More clothing => lower infrared reading Clo = k * (t clothing – t background ) + b t clothing is the infrared measured from clothes on human body t background is the background infrared radiation k and b are parameters to be estimated by regression
Personalization PMV model represents the average for a single office, only the occupant’s vote matters Predicted Personal Vote (PPV) Model ppv = f ppv (pmv) where f ppv () is a linear function
SPOT+: A Smart Personalized Office Thermal Control System Occupancy Prediction Learning-Based Modeling 500W f () + 1 o C -> Personal Thermal Comfort Evaluation Arrive officeLunch Setpoint Scheduling
Learning-Based Model Predictive Control We model the thermal characteristics of a room using LBMPC The model can predict future temperature = f lbmpc (current temperature, heater power)
Learning-Based Model Predictive Control
SPOT+: A Smart Personalized Office Thermal Control System Occupancy Prediction Learning-Based Modeling 500W f () + 1 o C -> Personal Thermal Comfort Evaluation Arrive officeLunch Setpoint Scheduling
Occupancy Prediction We predict occupancy using historical data. Match Previous similar history Predict using matched records _______________________________ [1] James Scott et. al., PreHeat: Controlling Home Heating With Occupancy Prediction, UbiComp 2011
SPOT+: A Smart Personalized Office Thermal Control System Occupancy Prediction Learning-Based Modeling 500W f () + 1 o C -> Personal Thermal Comfort Evaluation Arrive officeLunch Setpoint Scheduling
Optimal Control We use the optimal control strategy to schedule the setpoint over a day. The control objective is to reduce energy consumption and still maintain thermal comfort Overall energy consumption in the optimization horizon S Weight of comfort, set to large value to guarantee comfort first Predicted occupancy, we only guarantee comfort when occupied. aka m(s) = 1 Thermal comfort penalty. Both term equal zero when the user feels comfortable
Optimal Control - Constraints ε is the tolerance of predicted personal vote (PPV) So when | ppv(x(s)) | is smaller than ε, there is no penalty Otherwise, either β c (s) or β h (s) will be positive to penalize the discomfort thermal environment
Evaluation
Evaluation of clothing level estimation Root mean square error (RMSE) = Linear correlation =
Predicted Personal Vote Estimation Root mean square error (RMSE) = Linear correlation =
Accuracy of LBMPC The RMSE over a day is 0.17C.
Accuracy of Occupancy Prediction The result of optimal prediction is affected by occupancy prediction. False negative 10.4% (From 6am. to 8pm.) False positive 8.0% (From 6am. to 8pm.) Still an open problem
Comparison of schemes
Limitations SPOT+ requires thermal Insulation for personal thermal control Current SPOT+ costs about $1000 PPV requires some initial calibration State of window/door is not modelled in the current LBMPC Accuracy of clothing level estimation is affected by Accuracy of Kinect Distance effect of the infrared sensor
Conclusion We extended PMV model for personalized thermal control We design and implement SPOT+ We use LBMPC and optimal control for personalized thermal control SPOT+ can accurately maintain personal comfort despite environmental fluctuations allows a worker to balance personal comfort with energy use.
Relationship between PPV and Energy cost Maintaining a PPV of 0 consumes about 6 kWh electricity daily. By setting the target PPV to -0.5, we can save about 3 kWh electricity per day.
Average Discomfort vs Energy Consumption