Enabling Building Energy Auditing Using Adapted Occupancy Models 40 - - PowerPoint PPT Presentation

40 60 80 100 120 40 60 80 mm

Enabling Building Energy Auditing Using Adapted Occupancy Models

Ankur U. Kamthe, Varick L. Erickson, Miguel ´

• A. Carreira-Perpi˜

n´ an and Alberto E. Cerpa

{akamthe,verickson,mcarreira-perpinan,acerpa}@ucmerced.edu

BuildSys 2011

40 60 80 100 120 40 60 80 mm

Introduction Related Work Adapting Occupancy Models Performance Evaluation Discussion Summary

HVAC Systems

Heating, Ventilation and Air-Conditioning (HVAC) systems account for majority (≈50%) of building energy consumption (2008)∗.

◮ Assumption: Condition based on maximum room occupancy ◮ Rooms are often unoccupied or partially occupied ◮ Leads to inefficient environmental conditioning ◮ Optimize energy usage using systems that actuate using occupancy

models Alternatively, ensure that buildings adhere to the strictest energy efficiency standards.

∗ Source: Building Energy Data Book

(http://buildingsdatabook.eren.doe.gov/docs/htm/1.1.4.htm) 2 / 23

40 60 80 100 120 40 60 80 mm

Introduction Related Work Adapting Occupancy Models Performance Evaluation Discussion Summary

So where are all the green buildings?

Figure: South Hall - UC Berkeley (built 1873)

◮ Majority of existing buildings are older than 20 years. ◮ Do not meet current energy efficiency construction standards. ◮ Impact long-term energy consumption. ◮ Energy audits: energy savings through retrofitting.

3 / 23

40 60 80 100 120 40 60 80 mm

Introduction Related Work Adapting Occupancy Models Performance Evaluation Discussion Summary

Building Energy Auditing

◮ Involves inspection and analysis of the energy consumption from

utility bills.

◮ Deploy sensors on-site to measure and verify energy use. ◮ Onsite work takes 1-2 days. ◮ Data is input to DOE-2, EnergyPlus, etc. to evaluate and

recommend energy retrofits. Further, maximize energy savings by including occupancy model information within energy audits.

4 / 23

40 60 80 100 120 40 60 80 mm

Introduction Related Work Adapting Occupancy Models Performance Evaluation Discussion Summary

Occupancy model caveats

◮ Large training datasets (weeks, months). ◮ Models are specific to the building. ◮ Therefore, for all other buildings, again collect large training dataset.

5 / 23

40 60 80 100 120 40 60 80 mm

Introduction Related Work Adapting Occupancy Models Performance Evaluation Discussion Summary

Problem Statement

How can we maximize energy savings by using occupancy models in building energy audits when collecting only 1-2 days of occupancy traces?

6 / 23

40 60 80 100 120 40 60 80 mm

Introduction Related Work Adapting Occupancy Models Performance Evaluation Discussion Summary

Occupancy Modeling

◮ Modern buildings have submetering systems, electronic locking

systems, etc.

◮ Use of wireless sensor networks for other buildings:

◮ PIR sensors - binary indicators. ◮ Camera sensors - people counters.

◮ Data collection goal: collect data for 1-2 days.

7 / 23

40 60 80 100 120 40 60 80 mm

Introduction Related Work Adapting Occupancy Models Performance Evaluation Discussion Summary

Proposed Approach

Use a reference building occupancy model that has been trained with extensive data and adapt it to the new building given a far smaller

• ccupancy data trace than would be necessary to train a new model

from scratch.

8 / 23

40 60 80 100 120 40 60 80 mm

Introduction Related Work Adapting Occupancy Models Performance Evaluation Discussion Summary

Reference Model

◮ Mixture of multivariate Gaussians (M) components in place of a

single multivariate Gaussian for every hour.

◮ Parameters: means (µ) and covariance matrix (Σ) for each hour. ◮ (D + 1)M + D2 − 1 parameters for every hour (D = #rooms). ◮ Use Expectation-Maximization (EM) algorithm for parameter

estimation (Retraining).

9 / 23

40 60 80 100 120 40 60 80 mm

Introduction Related Work Adapting Occupancy Models Performance Evaluation Discussion Summary

Reference and Adaptation Model Datasets

1 1 1 2

Student Lab Conference Room Hallway Office

(a) (b) Reference Adaptation

10 / 23

40 60 80 100 120 40 60 80 mm

Introduction Related Work Adapting Occupancy Models Performance Evaluation Discussion Summary

Reference and Adaptation Model Datasets

6 12 18 24 1 2 3 Hour of Day

• Num. of Occupants

Hall 1 6 12 18 24 1 2 3 4 Hour of Day Hall 2 6 12 18 24 5 10 Hour of Day Office 1 6 12 18 24 5 10 Hour of Day Lab 1

(a) Reference Dataset Room Occupancy

6 12 18 24 1 2 3 Hour of Day

• Num. of Occupants

Hall 6 12 18 24 1 2 3 4 Hour of Day Conference 6 12 18 24 5 10 Hour of Day Office 6 12 18 24 5 10 Hour of Day Lab

(b) Adaptation Dataset Room Occupancy

Figure: Room occupancy averaged over the length of dataset (5-days) for every hour for the reference model (a) and adaptation (and retrained) model (b), respectively.

11 / 23

40 60 80 100 120 40 60 80 mm

Introduction Related Work Adapting Occupancy Models Performance Evaluation Discussion Summary

Adapted Model

◮ Assumptions: Well-trained reference model and occupancy data for

target (audited) building.

◮ Adaptation Approach: Tie the means of the reference model

using a non-linear transformation.

12 / 23

40 60 80 100 120 40 60 80 mm

Introduction Related Work Adapting Occupancy Models Performance Evaluation Discussion Summary

Adaptation Illustration with 2 Rooms

4 6 8 10 12 20 40 60

Ref Data Histogram

9 10 11

Comp 1 µ1 = 10

4 5 6

Comp 2 µ2 = 5

8 10 12 1 2 3 4 5

Adapt Data Histogram

13 / 23

40 60 80 100 120 40 60 80 mm

Introduction Related Work Adapting Occupancy Models Performance Evaluation Discussion Summary

Adaptation Illustration with 2 Rooms

4 6 8 10 12 20 40 60

Ref Data Histogram

9 10 11

Comp 1 µ1 = 10

4 5 6

Comp 2 µ2 = 5

• µd =

Om 1+e−(aµd /Om+b)

a = −3.1183 b = 3.6972

• Ref. Parms.
• Adapt. Parms.

5 12 10 12 9 12 11 12

Sigmoid Transformation

Om = 12 8 10 12 1 2 3 4 5

Adapt Data Histogram

8 9 10

Comp 1 µ1 = 9

10 11 12

Comp 2 µ2 = 11

14 / 23

40 60 80 100 120 40 60 80 mm

Introduction Related Work Adapting Occupancy Models Performance Evaluation Discussion Summary

Adapted vs Retrained Model

◮ The objective function is the log-likelihood of the adaptation data

given the constrained MVGM with 3M − 1 free parameters: L

• {

πm, am, bm}M

m=1

• = N

n=1 log M m=1

πmp(xn; am, bm)

◮ Adaptation: 3M − 1 adaptation parameters. ◮ Retraining objective function

L

• {

πm, µhm, Σh}M

m=1

• = N

n=1 log M m=1

πmp(xn; µhm, Σh)

◮ Retraining: (D + 1)M + D2 − 1 parameters for every hour.

15 / 23

40 60 80 100 120 40 60 80 mm

Introduction Related Work Adapting Occupancy Models Performance Evaluation Discussion Summary

Modeling Performance

1 2 3 4 −4 −3 −2 −1 x 10

8

Adaptation Dataset (in days) Loglikelihood (on test set)

Adaptation Retraining Optimal

Figure: Log-likelihood of the different models as a function of the days in the adaptation dataset.

16 / 23

40 60 80 100 120 40 60 80 mm

Introduction Related Work Adapting Occupancy Models Performance Evaluation Discussion Summary

Estimated occupancy models

6 12 18 24 5 10

Hour of Day

• Num. of Occupants

Office

4DayRetrain 1DayRetrain 1DayAdapt

17 / 23

40 60 80 100 120 40 60 80 mm

Introduction Related Work Adapting Occupancy Models Performance Evaluation Discussion Summary

Estimated occupancy models

6 12 18 24 1 2 3 4 5

Hour of Day

Conference Room

4DayRetrain 1DayRetrain 1DayAdapt

18 / 23

40 60 80 100 120 40 60 80 mm

Introduction Related Work Adapting Occupancy Models Performance Evaluation Discussion Summary

Building Energy Simulation Results

◮ Construct occupancy schedule using models 4DayRetrain

(MVGM-R4), 1DayRetrain (MVGM-R1) and 1DayAdapt (MVGM-A1).

◮ EnergyPlus model of the building floorplan (total 32,000 sq.ft.)

from which we have adaptation data for a Hall, Office, Lab and Conference room (approx. 12,000 sq.ft.)

◮ Compare to:

◮ Baseline: maximum room occupancy between 7a.m.-10p.m. and is

• ff at other times.

◮ OBSERVE: Markov chain approach to model the temporal changes in

• ccupancy of a building. Close to optimal conditioning.

19 / 23

40 60 80 100 120 40 60 80 mm

Introduction Related Work Adapting Occupancy Models Performance Evaluation Discussion Summary

Energy Savings

Apr Aug Dec 0.05 0.1 0.15 0.2 0.25

Month of Year Energy Savings (%)

OBSERVE 4DayRetrain 1DayAdapt 1DayRetrain

1DayAdapt (10.9%) < OBSERVE (11.2%) < 1DayAdapt (11.4%) < 1DayRetrain (12.9%)

20 / 23

40 60 80 100 120 40 60 80 mm

Introduction Related Work Adapting Occupancy Models Performance Evaluation Discussion Summary

Conditioning Effectiveness

0:00 4:00 8:00 12:00 16:00 20:00 24:00 4 8 12 Hour RMSE

RMSE Target Temperature: Conference Room, Winter

1DayAdapt 4DayRetrain OBSERVE Baseline 1DayRetrain

0:00 4:00 8:00 12:00 16:00 20:00 24:00 4 8 12 Hour RMSE

RMSE Target Temperature: Conference Room, Summer

1DayAdapt 4DayRetrain OBSERVE Baseline 1DayRetrain

Summer: 4DayRetrain, OBSERVE, 1DayAdapt (< 0.5oF) < 1DayRetrain (1.8oF) Winter: 4DayRetrain, OBSERVE, 1DayAdapt (< 1.4oF) < 1DayRetrain (2.4oF)

21 / 23

40 60 80 100 120 40 60 80 mm

Introduction Related Work Adapting Occupancy Models Performance Evaluation Discussion Summary

Lessons Learned

◮ Retraining with little data leads to poor models. ◮ Adapted model generalizes well ..... if the reference model is close

enough to the new adaptation occupancy data.

22 / 23

40 60 80 100 120 40 60 80 mm

Introduction Related Work Adapting Occupancy Models Performance Evaluation Discussion Summary

Summary

◮ Use model adaptation for construction of good-quality occupancy

models with 1-day of occupancy data.

◮ Conditioning effectiveness on par with other models that require 4

times as much training data.

◮ Energy auditing improvements using adapted occupancy models.

Thank You .... Questions?

23 / 23