buildsys15 - Speaker Deck

Slide 1

Slide 1 text

Nipun Batra, Amarjeet Singh, Kamin Whitehouse Buildsys 2015 If You Measure It, Can You Improve It? Exploring The Value of Energy Disaggregation through actionable feedback 1

Slide 2

Slide 2 text

General eco feedback vs Actionable Feedback Eco feedback Misc. 22% Light 10% Fridge 11% HVAC 56%

Slide 3

Slide 3 text

General eco feedback vs Actionable Feedback Eco feedback Misc. 22% Light 10% Fridge 11% HVAC 56%

Slide 4

Slide 4 text

General eco feedback vs Actionable Feedback Eco feedback Misc. 22% Light 10% Fridge 11% HVAC 56% Power (W) 0 175 350 525 700 Power (W) 0 175 350 525 700 Home 1 Home 2 Actionable feedback Fridge consumption over 24 hours Misc. 22% Light 10% Fridge 11% HVAC 56%

Slide 5

Slide 5 text

Slide 6

Slide 6 text

Slide 7

Slide 7 text

General eco feedback vs Actionable Feedback Eco feedback Misc. 22% Light 10% Fridge 11% HVAC 56% Power (W) 0 175 350 525 700 Home 2 Actionable feedback Fridge consumption over 24 hours Your fridge defrosts too much, wasting 30% energy Misc. 22% Light 10% Fridge 11% HVAC 56%

Slide 8

Slide 8 text

Approach overview- How to give feedback Power (W) 0 175 350 525 700 Speciﬁc features of trace to infer why energy usage is high Length of duty cycles

Slide 9

Slide 9 text

Approach overview- How to give feedback Power (W) 0 175 350 525 700 Speciﬁc features of trace to infer why energy usage is high Actual power value

Slide 10

Slide 10 text

Feedback methods on Fridge and HVAC Both appliances commonly found across homes Others 38% Fridge 8% HVAC 54%

Slide 11

Slide 11 text

Evaluation overview Submetered traces Power (W) 0 350 700 Submeter sensor

Slide 12

Slide 12 text

Can we give such feedback? Disaggregated traces Power (W) 0 350 700 NILM Household aggregate Submetered traces Power (W) 0 350 700 Submeter sensor 0 2000 4000 0 2000 4000 Smart meter

Slide 13

Slide 13 text

Do disaggregated traces provide features needed for providing feedback? Disaggregated traces Power (W) 0 350 700 NILM Household aggregate Submetered traces Power (W) 0 350 700 Submeter sensor 0 2000 4000 0 2000 4000 Smart meter

Slide 14

Slide 14 text

0 125 250 375 500 Fridge is a duty cycle based appliance; compressor turns ON and OFF periodically

Slide 15

Slide 15 text

Defrost cycles occurs periodically and consume high amount of power 0 125 250 375 500

Slide 16

Slide 16 text

0 125 250 375 500 Defrost introduces heat increasing ON duration of next cycles

Slide 17

Slide 17 text

Fridge usage increases compressor ON durations (and reduce compressor OFF durations) 0 125 250 375 500

Slide 18

Slide 18 text

Night hours typically have “baseline” usage 0 175 350 525 700 Baseline duty % = Median duty % in the night

Slide 19

Slide 19 text

Defrost energy 0 175 350 525 700 Defrost energy = Energy consumed in defrost state + Extra energy consumed in next few compressor cycles

Slide 20

Slide 20 text

Defrost energy 0 175 350 525 700 Defrost energy = Energy consumed in defrost state + Extra energy consumed in next few compressor cycles

Slide 21

Slide 21 text

Usage energy 0 175 350 525 700 Usage energy = Extra energy consumed over baseline

Slide 22

Slide 22 text

Experimental setup Wiki Energy data set 1. 97 fridges 2. 58 HVAC

Slide 23

Slide 23 text

13 out of 95 homes can be given feedback based on usage energy saving upto 23% fridge energy

Slide 24

Slide 24 text

13 out of 95 homes can be given feedback based on usage energy saving upto 23% fridge energy NILM algorithms show poor accuracy in identifying homes which can be given feedback based on usage energy

Slide 25

Slide 25 text

17 out of 95 homes can be given feedback on excess defrost saving upto 25% fridge energy

Slide 26

Slide 26 text

Such feedback can’t be given with disaggregated traces, since these techniques fare poorly on defrost detection.

Slide 27

Slide 27 text

Benchmark NILM algorithms on our data set give accuracy comparable or better than state-of-the-art Kolter 2012 Parson 2012 Parson 2014 Batra 2014 CO FHMM Hart Error Energy % 0 17.5 35 52.5 70

Slide 28

Slide 28 text

0 125 250 375 500 “Average” error in energy would be low even if NILM predicted this

Slide 29

Slide 29 text

But, we wanted to predict.. 0 125 250 375 500

Slide 30

Slide 30 text

It’s the details that we care about 0 125 250 375 500

Slide 31

Slide 31 text

Like fridge, HVAC duty cycles to maintain the set temperature 0 1000 2000 3000 4000

Slide 32

Slide 32 text

As temperature increases during the day, more energy required to cool the home 0 1000 2000 3000 4000

Slide 33

Slide 33 text

0 1000 2000 3000 4000 People typically turn up the temperatures when they leave home

Slide 34

Slide 34 text

Recommended Min. Temp (F) 77 79 81 83 85 2 4 6 8 10 12 14 16 18 20 22 24 EnergyStar.gov recommended HVAC setpoint schedule Sleep Morning Work Evening

Slide 35

Slide 35 text

Recommended Min. Temp (F) 77 79 81 83 85 2 4 6 8 10 12 14 16 18 20 22 24 Setpoint schedule score Sleep Morning Work Evening

Slide 36

Slide 36 text

Recommended Min. Temp (F) 77 79 81 83 85 2 4 6 8 10 12 14 16 18 20 22 24 Setpoint schedule score Sleep Sleep score = 1 if sleep temp. > 82, (82-temp.)/4 if 78

Slide 37

Slide 37 text

Learning HVAC setpoint 0 1000 2000 3000 4000 77 85 5 1015 20 HVAC trace Weather Learnt setpoint

Slide 38

Slide 38 text

Giving feedback 77 85 5 1015 20 Features from HVAC trace 69 75 5 10 15 20 0 1000 2000 3000 4000 77 85 5 10 15 20 Learnt setpoint Don’t need feedback Need feedback

Slide 39

Slide 39 text

84% accuracy on giving feedback using submetered traces 39

Slide 40

Slide 40 text

NILM methods give 15-30% worse accuracy for feedback 40

Slide 41

Slide 41 text

Benchmark NILM algorithms on our data set give accuracy comparable or better than state-of-the-art Batra 2014 CO FHMM Hart Error Energy % 0 7.5 15 22.5 30

Slide 42

Slide 42 text

Error in prediction of minutes of HVAC usage (%) 0 6 12 18 24 Hart FHMM CO Night Morning Morning hours which have lesser NILM accuracy are important for HVAC feedback

Slide 43

Slide 43 text

Conclusions Appliance level data does enable actionable energy saving feedback

Slide 44

Slide 44 text

Conclusions Appliance level data does enable actionable energy saving feedback BUT Results show that we need to revisit the metrics by which we measures progress