sensys_dc15 - Speaker Deck

Slide 1

Slide 1 text

Nipun Batra IIIT Delhi November 1, 2015 Making energy disaggregation practical

Slide 2

Slide 2 text

Buildings contribute signiﬁcantly to overall energy consumption Percentage contribution to overall energy 0 12.5 25 37.5 50 Australia China India Korea USA

Slide 3

Slide 3 text

Buildings getting constructed at rapid rate

Slide 4

Slide 4 text

Buildings are an attractive target towards sustainability

Slide 5

Slide 5 text

Residential buildings can contribute upto 93% of building energy usage

Slide 6

Slide 6 text

“If you cannot measure it, you cannot improve it” - Kelvin

Slide 7

Slide 7 text

Sensor deployments have several challenges* 1. Homes are not a power panacea 2. Homes have poor connectivity 3. Homes are hazardous 4. Limited user interaction 5. Aesthetics matter 1.Hnat et al. “The hitchhiker's guide to successful residential sensing deployments”. Sensys 2010 2.Batra et al. “It’s different. Insights into home energy consumption in India”. Buildsys 2013

Slide 8

Slide 8 text

Non intrusive load monitoring (NILM) or Energy disaggregation Smart meter Machine learning

Slide 9

Slide 9 text

Why NILM can work Different appliances can have unique “signatures”

Slide 10

Slide 10 text

Making NILM practical 1.Comparable 2.Utility-driven 3.Scalable

Slide 11

Slide 11 text

Making NILM comparable eEnergy 2014 and Buildsys 2014

Slide 12

Slide 12 text

What is the best NILM approach? Despite 30+ years of NILM research really hard question 3 main problems

Slide 13

Slide 13 text

1: Hard to assess generality • Previous contributions evaluated only on single dataset • Non-trivial to set up similar experimental conditions for direct comparison

Slide 14

Slide 14 text

2: Lack of comparison against same benchmarks • Newly proposed algorithms rarely compared against same benchmarks • Lack of “open source” reference algorithms often lead to reimplementation

Slide 15

Slide 15 text

3: “Inconsistent” disaggregation performance metrics • Different performance metrics proposed in the past • Different formulae for same metric, eg. 4+ versions of “energy assigned”

Slide 16

Slide 16 text

And NILMTK was born Open source NILM toolkit to enable easy comparative analysis of NILM algorithms across data sets

Slide 17

Slide 17 text

NILMTK pipeline REDD BLUED UK- DALE Statistics NILMTK- DF Training Preprocessing Model Disaggregation Metrics Data interface

Slide 18

Slide 18 text

NILMTK-DF: Common data format REDD BLUED UK- DALE Statistics NILMTK- DF Training Preprocessing Model Disaggregation Metrics Data interface 10 data sets released

Slide 19

Slide 19 text

Statistical functions REDD BLUED UK- DALE Statistics NILMTK- DF Training Preprocessing Model Disaggregation Metrics Data interface Suite of commonly used statistical functions Ground truth quality 0 50 100 REDD SMART* PECAN AMPds iAWE UK_DALE

Slide 20

Slide 20 text

Preprocessing REDD BLUED UK- DALE Statistics NILMTK- DF Training Preprocessing Model Disaggregation Metrics Data interface 18/04/11 06/05/11 24/05/11 Time (day/month/year) Fridge Washer dryer Kitchen outlets Mains 1 Mains 2 0 10 20 Dropout rate (%)

Slide 21

Slide 21 text

Train and Disaggregate REDD BLUED UK- DALE Statistics NILMTK- DF Training Preprocessing Model Disaggregation Metrics Data interface

Slide 22

Slide 22 text

Train and Disaggregate Hart’s event detection algorithm Factorial Hidden Markov Model (FHMM) Appliance Off power On power Light 0 200 Fridge 0 100 Combinatorial Optimisation

Slide 23

Slide 23 text

NILMTK impact • 10+ papers using NILMTK (4 in Buildsys 2015) • 2 user contributed NILM algorithms • 3 user contributed NILM data sets • Best demonstration award at Buildsys 2014

Slide 24

Slide 24 text

Making NILM utility-driven Buildsys 2015 and Percom 2016 (under submission)

Slide 25

Slide 25 text

If you can measure, can you improve?

Slide 26

Slide 26 text

Does NILM really save energy? Does telling you that HVAC takes 56% save you energy? Lack of speciﬁc actionable insights Misc. 22% Light 10% Fridge 11% HVAC 56%

Slide 27

Slide 27 text

Exploring the value of Energy disaggregation 1. Can disaggregated traces provide actionable insights? Submetered appliance traces Appliance energy models Appliance energy modelling Identify homes needing feedback

Slide 28

Slide 28 text

Exploring the value of Energy disaggregation 2. Do existing NILM techniques provide traces with sufﬁcient ﬁdelity to support feedback? Disaggregated appliance traces Appliance energy models Appliance energy modelling Identify homes needing feedback Aggregate household power NILM

Slide 29

Slide 29 text

Feedback methods on Fridge and HVAC • Both appliances common across homes • Both appliances contribute heavily to overall energy consumption

Slide 30

Slide 30 text

Fridge energy modelling

Slide 31

Slide 31 text

We can break down fridge energy with less than 4% error

Slide 32

Slide 32 text

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

Slide 33

Slide 33 text

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

Slide 34

Slide 34 text

HVAC modelling Objective 1. Learn set point from weather and energy data 2. Optimising setpoint can save upto 20-30% HVAC

Slide 35

Slide 35 text

HVAC feedback • 84% accuracy on giving feedback based on setpoint temperature

Slide 36

Slide 36 text

Slide 37

Slide 37 text

Benchmark NILM algorithms on our data set give accuracy comparable to state-of-the-art

Slide 38

Slide 38 text

Slide 39

Slide 39 text

NILM algorithms show poor accuracy in identifying homes which can be given feedback based on usage energy

Slide 40

Slide 40 text

NILM algorithms don’t identify the defrost state and thus prevent feedback based on defrost energy Defrost state is hard to detect!

Slide 41

Slide 41 text

NILM algorithms show poor accuracy in identifying homes needing HVAC setpoint feedback

Slide 42

Slide 42 text

Take aways 1. Appliance level data does enable actionable energy saving feedback 2. Feedback accuracy can be low despite good disaggregation accuracy 3. We, the disaggregation community, need to revisit the metrics by which we measure progress

Slide 43

Slide 43 text

Making NILM scalable IPSN 2016 (under submission)

Slide 44

Slide 44 text

3 fundamental problems 1. Lights (and other low power appliances) show poor disaggregation accuracy. Light are third highest overall in terms of loads 2. Current NILM algorithms are often supervised and need careful tuning and model speciﬁcation. 3. Most techniques assume 1 min. or less sampling interval. Existing smart meters sample once every 15 mins.

Slide 45

Slide 45 text

Can we leverage big data to make NILM more scalable?

Slide 46

Slide 46 text

Is big data more valuable than precise data?

Slide 47

Slide 47 text

precise data • Smart meter • 1 min sampling • Fine tune model per home

Slide 48

Slide 48 text

big data • Large number of homes

Slide 49

Slide 49 text

big data • Large number of homes • Submeter small subset of homes • Use single reading per month

Slide 50

Slide 50 text

Similar homes have similar per- appliance energy consumption

Slide 51

Slide 51 text

Approach: Neighbourhood NILM

Slide 52

Slide 52 text

Features • Energy consumption: • Past 12 months household aggregate • Ratios (Min. energy/Max. energy) • Static household properties: #occupants, Area, #rooms

Slide 53

Slide 53 text

Neighbourhood NILM comparable or better than best reported NILM accuracy Fridge HVAC Washing machine Dish washer Dryer Lights 0 20 40 60 80 100 Energy accuracy (%) (Higher is better) National average FHMM Neighbourhood NILM Best reported NILM accuracy Oracle

Slide 54

Slide 54 text

Neighbourhood NILM signiﬁcantly accurate in Washing machine, dish washer, dryer- all pain points for traditional NILM Fridge HVAC Washing machine Dish washer Dryer Lights 0 20 40 60 80 100 Energy accuracy (%) (Higher is better) National average FHMM Neighbourhood NILM Best reported NILM accuracy Oracle

Slide 55

Slide 55 text

High accuracy of “Oracle” suggests that our approach is promising Fridge HVAC Washing machine Dish washer Dryer Lights 0 20 40 60 80 100 Energy accuracy (%) (Higher is better) National average FHMM Neighbourhood NILM Best reported NILM accuracy Oracle

Slide 56

Slide 56 text

Take away Big data more valuable than precise data for the problem of energy disaggregation

Slide 57

Slide 57 text

Conclusions 1.Comparable- NILMTK 2.Utility-driven-Energy saving feedback Inferring household characteristics 3.Scalable- Neighbourhood NILM Making NILM practical in 3 ways:

Slide 58

Slide 58 text

Future work

Slide 59

Slide 59 text

Neighbourhood NILM with 15 minute meter data 1. Can we reduce the number of neighbours needed when we use 15 minute meter data 2. 15 minute data will present daily patterns, in addition to monthly patterns in current implementation 3. Metrics and utilities on 15 minute resolution

Slide 60

Slide 60 text

Homes “changing” behaviour pose an interesting challenge to Neighbourhood NILM 1. Balance between “historical” data and recent trends? 2. Continue having same neighbours? 1. When to “change” the neighbours of a home

Slide 61

Slide 61 text

Scaling NILM to “similar” commercial buildings/different appliance types 1. Class of commercial buildings have exact same electrical infrastructure 2. Deployment across 10 dairy booths in New Delhi

Slide 62

Slide 62 text

NILMTK-“The cost of impact is a bug report/feature request a day on Github” :)

Slide 63

Slide 63 text

Conclusions 1.Comparable- NILMTK 2.Utility-driven-Energy saving feedback Inferring household characteristics 3.Scalable- Neighbourhood NILM Making NILM practical in 3 ways:

Slide 64

Slide 64 text

Other work 1. Insights into home energy consumption in India [Buildsys 2013] 2. Inferring household characteristics from NILM [under submission Percom 2016] 3. Improving NILM performance using additional data [ICMLA 2013]