Thermal Modeling for a HVAC Controlled Real-life Auditorium Chenyang - - PowerPoint PPT Presentation

Thermal Modeling for a HVAC Controlled Real-life Auditorium

Chenyang Lu

Endeavor on Smart Building

• 1. Instrumenting an auditorium
• 2. Modeling spatiotemporal thermal dynamics
• 3. Occupancy-based energy saving for HVAC
• 4. Micro-metering an apartment

Challenges

Ø Heat, Ventilation and Air Conditioning (HVAC) consumes 33% of building energy. Ø HVAC control relies on accurate thermal models. Ø Large open spaces have complex spatiotemporal dynamics.

q Examples: auditoriums, theatres, open offices, lobbies.

3

Week-long temperature trace at different locations in an auditorium.

Spa6al Varia6on in an Auditorium

4

• Temperature differs by ~2°C despite HVAC control.
• Unique challenges in large open spaces.

Experimental Approach

• 1. Deploy 34 sensors in an auditorium for over three months.
• 2. Collect multimodal data to capture fine-grained

spatiotemporal dynamics under HVAC control .

• 3. Identify thermal model based on data from all sensors.
• 4. Simplify model through sensor selection.

5

Instrumen6ng an Auditorium

Ø Emerson wireless sensors: temperature, humidity. Ø HVAC sensors: air flow rate and temperature. Ø Wireless camera: occupancy and lighting (on/off).

6

Thermostats Camera Wireless Sensors

Brauer Hall 1/2013 - 5/2013

Wireless Monitoring System

Base Station

Particle Sensor Temperature Sensor Temperature Sensor

Brauer Hall Database

W i r e l e s s L i n k s Wireless Links Wireless Links W i r e l e s s L i n k s

Auditorium

CO2 Sensor CO2 Sensor

Data Analysis Empirical Study

Humidity Sensor Temperature Sensor

surveillance camera

Occupancy

7

Instrumen6ng the Auditorium

Ø Environmental monitoring

q 34 temperature sensors q 15 humidity sensors q 1 condensa;on par;cle counter q 2 CO2 sensors

Ø HVAC: air flow rate, air temperature Ø Occupancy from camera Ø Database

q Sensors con;nuously feed data to database over the Internet q Visualiza;on through web interface

8

Large Mul6-modal Dataset

Ø Longitude: >8 months of data Ø Fine grained

q Temperature: 1 reading per 1/3 degree change q Humidity: 1 reading per 1% degree change q Particle: 3 readings/second q CO2: 2 readings/hour q HVAC air flow: 4 readings/hour q Occupancy: 4 photos/hour

9

Temperature & Humidity Sensor

10

Ø Emerson wireless thermostats

q

Repurposed for distributed monitoring

Ø Capture fine-grained spatiotemporal dynamics

q

Improve HVAC model and control 2/25/13 – 3/3/13

Wireless Condensa6on Par6cle Counter

11

butanol particle number concentration (µgm-3) display inlet

wireless transmitter Instrument specifica6ons

• Uses butanol, single-count, and

photometric technology, to count airborne par;cle numbers with a diameter from 0.07 to 3 µm

• Fast response ;me (<13 seconds)
• Semi-portable
• High-resolu;on (1Hz) data

Particle sources

people furnishings (chairs, carpet) hot food

• utdoors

(traffic, dust) HVAC resuspension

• RetrofiHed with

Bluetooth.

• Help understand

impacts of HVAC on air quality

Par6cle Number Concentra6ons

12 HVAC switching to

• ff mode (~ 9pm)

Sunday

Midterm exam

One Week of Data Traces (2/25 – 3/3)

13 13

Temperature Occupancy CO2 Particle Air flow

Endeavor on Smart Building

• 1. Instrumenting an auditorium
• 2. Modeling spatiotemporal thermal dynamics
• 3. Occupancy-based energy saving for HVAC
• 4. Micro-metering an apartment

Prior Modeling Approaches

Ø Principle-driven: rely on detailed knowledge of building design and materials. Ø Data-driven: estimate model based on data.

q Assume same temperature per room: ignore spatial variations and

interactions within a large space.

q Divide space into zones: reply on known inter-zone interactions.

15

Model Iden6fica6on

Ø Model identification based on training data

q Minimize modeling error with least square optimization q Solved using CVX toolbox for Matlab

Ø Tradeoff between model complexity and accuracy

q 1st order model à simple q 2nd order model à capture more complex dynamics

16

T(k+1) = AT(k) + BU(k) Temperature T(k) U(k): air flow rate & temperature,

• ccupancy, light.

Estimated temperature T(k+1)

1st vs. 2nd Order Model

Ø 2nd order model more accurately captures the spatiotemporal dynamics in the auditorium.

17

Measured vs. predicted temperature on 2/28/13

Model Simplifica6on

Ø Disadvantages of fine-grained models based on all sensors

q Complex model is unsuitable for control design. q Challenge in maintaining numerous sensors.

Ø Approach: simplifying model through sensor selection

q Sensor data have strong correlations. q Select a subset of sensors to capture spatiotemporal dynamics. q Identify thermal model based on selected sensors.

Ø Advantage of model simplification

q Practical for HVAC control. q Only need to keep the selected sensors during operation. q Dense sensor network needed only initially to collect training data.

18

Sensor Selec6on based on Clustering

• 1. Spectral clustering based on sensor data.

q Value: group sensors with similar temperature values. q Correlation: group sensors whose data traces follow similar trends.

19

Correla6on-based Sensor Clustering

20

Two clusters Temperature correlation

Sensor Selec6on based on Clustering

• 1. Spectral clustering based on sensor data.

q Value: group sensors with similar temperature values. q Correlation: group sensors whose data traces follow similar trends.

• 2. Select a sensor from each cluster.

q Stratified Random Selection (SRS): randomly choose one. q Stratified Mean Selection (SMS): select the sensor whose data is the

closest to the cluster mean.

21

Model Simplifica6on

22

• Clustering outperforms Random Selection (RS)
• Stratified Mean Selection (SMS) is more accurate than Stratified

Random Selection, especially for large clusters.

Summary: Thermal Modeling

Ø Large open spaces have complex spatiotemporal dynamics. Ø Data-driven thermal modeling for large open spaces.

1.

Sensor network captures spatiotemporal dynamics.

2.

Sensor selection based on data clustering.

3.

Model identification based on data of selected sensors.

Ø Validated on data collected from a real-life auditorium. Ø Exciting opportunities ahead

q Optimize HVAC control q Leverage air quality sensing for more aggressive energy saving

23

• Y. Fu, M. Sha, C. Wu, A. KuYa, A. Leavey, C. Lu, H. Gonzalez, W. Wang, B. Drake, Y. Chen and P. Biswas, Thermal

Modeling for a HVAC Controlled Real-life Auditorium, ICDCS 2014.

Endeavor on Smart Building

• 1. Instrumenting an auditorium
• 2. Modeling spatiotemporal thermal dynamics
• 3. Occupancy-based energy saving for HVAC
• 4. Micro-metering an apartment

HVAC Energy Waste

Ø Current HVAC operates on fixed schedule

q On (occupied mode) during daytime (6am-9pm) q Off (non-occupied mode) at night

Ø But the auditorium is vacant 80% of the time during the day!

25

Note: Occupancy Follows Calendar

Ø Calendar predicts actual occupancy at >98% accuracy

q Validated by camera Seminar Class Meeting

26

Ø Preconditioning: Start HVAC Tp before an event

q Tp: time needed to reach the temperature set point q Tp = 3 hours for the auditorium based on data traces

Ø Save energy: Turn off HVAC if >Tp till next event

q Turn off HVAC immediately after the last event each day q HVAC remains off during weekends

Ø Avoid thrashing: remains on if next event is within Tp

q Maintain comfort q Reduce unnecessary switching

Schedule HVAC based on Calendar

27

Example

Turn off aaer last event Precondi;oning 3 hours Interval between events less than 3 hours On Off Sun Sat

28

q Turning off HVAC immediately after last event à 36% q Turning off HVAC on Sat/Sun à 34% q Turning on HVAC late in the morning à 8%

78% Energy Saving over 6 Weeks

29

Endeavor on Smart Building

• 1. Instrumenting an auditorium
• 2. Modeling spatiotemporal thermal dynamics
• 3. Occupancy-based energy saving for HVAC
• 4. Micro-metering an apartment

Smart Home: Objec6ves

Ø Save energy while maintaining comfort. Ø Close the loop: intelligent control of appliances. Ø Human centered: incentivize residents to save energy. Ø Internet of Things: integrate sensors, appliances, cloud, and smartphones.

31

The Internet of Things

Weather station AC Power meter BT / BTL listener 15.4 WiFi microserver

32

UI

Pilot – Components

Ø ACme – Berkeley power meter

q Based on the Epic core q Runs TinyOS q IPv6 over mesh network

Ø Raspberry Pi

q Very popular microserver

Ø Ethernet connection to apartment router Ø Amazon EC2 as the cloud Ø Measuring major appliances power consumption

Power meter microserver

33

Endeavor on Smart Building

• 1. Instrumenting an auditorium
• 2. Modeling spatiotemporal thermal dynamics
• 3. Occupancy-based energy saving for HVAC
• 4. Micro-metering an apartment