data science for energy efficiency (dmytro mindra technology stream)

71
Copyright © 2016, Oracle and/or its affiliates. All rights reserved. | Dmytro Mindra Senior Manager, ApplicaCons Engineering UCliCes Global Business Unit October 1, 2016 Data Science for Energy Efficiency

Upload: lviv-it-arena

Post on 15-Apr-2017

92 views

Category:

Technology


1 download

TRANSCRIPT

Copyright©2016,Oracleand/oritsaffiliates.Allrightsreserved.|Copyright©2016,Oracleand/oritsaffiliates.Allrightsreserved.|

DmytroMindraSeniorManager,ApplicaConsEngineeringUCliCesGlobalBusinessUnitOctober1,2016

Data Science for Energy Efficiency

Copyright©2016,Oracleand/oritsaffiliates.Allrightsreserved.|

Our Mission WeaimtoprovidetheindustrywiththemostcompletecloudplaLormfortheenCreuClityvaluechain,frommetertogridtoend-customers.

Network Meter-to-cash Endconsumer

Copyright©2016,Oracleand/oritsaffiliates.Allrightsreserved.|

over

50%of US residential

energy data

600+

billionmeter reads

nearly

2/3of US residential smart meter data

Copyright©2016,Oracleand/oritsaffiliates.Allrightsreserved.|

Opower’scustomerengagementplaLormcombinesinsighLulanalyCcs,behavioralscience,andcuPng-edgeUXtohelpuCliCes

elevatethecustomerexperience

Copyright©2016,Oracleand/oritsaffiliates.Allrightsreserved.|

• Estimate an unknown value

– Predict future usage

What is Machine Learning algorithms that solve a problem by learning from data

Copyright©2016,Oracleand/oritsaffiliates.Allrightsreserved.|

• Estimate an unknown value

– Predict future usage

– Estimate something about a home

What is Machine Learning

sqft

algorithms that solve a problem by learning from data

Copyright©2016,Oracleand/oritsaffiliates.Allrightsreserved.|

• Estimate an unknown value

– Predict future usage

– Estimate something about a home • Find patterns in data

What is Machine Learning algorithms that solve a problem by learning from data

Copyright©2016,Oracleand/oritsaffiliates.Allrightsreserved.|

• Want to estimate some value: – Does this household use GAS or ELECTRIC heat?

Standard machine learning setting

Copyright©2016,Oracleand/oritsaffiliates.Allrightsreserved.|

• Want to estimate some value: – Does this household use GAS or ELECTRIC heat?

»  Have something we know about each household that might help us estimate the unknown value

Standard machine learning setting

Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec

Copyright©2016,Oracleand/oritsaffiliates.Allrightsreserved.|

Estimating heat type

What do we know about a household that might help us estimate whether it has gas or electric heat?

Copyright©2016,Oracleand/oritsaffiliates.Allrightsreserved.|

Estimating heat type

0 10 20 30 40

Jan

Feb

Mar

A

pr

May

Ju

n Ju

l A

ug

Sep

O

ct

Nov

D

ec

kWh

0

5

10

Jan

Feb

Mar

A

pr

May

Ju

n Ju

l A

ug

Sep

O

ct

Nov

D

ec

Therms

Copyright©2016,Oracleand/oritsaffiliates.Allrightsreserved.|

Estimating heat type

0 10 20 30 40

Jan

Feb

Mar

A

pr

May

Ju

n Ju

l A

ug

Sep

O

ct

Nov

D

ec

kWh

0

5

10

Jan

Feb

Mar

A

pr

May

Ju

n Ju

l A

ug

Sep

O

ct

Nov

D

ec

Therms

Copyright©2016,Oracleand/oritsaffiliates.Allrightsreserved.|

Estimating heat type

0

5

10

Jan

Feb

Mar

A

pr

May

Ju

n Ju

l A

ug

Sep

O

ct

Nov

D

ec

Therms

0 10 20 30 40

Jan

Feb

Mar

A

pr

May

Ju

n Ju

l A

ug

Sep

O

ct

Nov

D

ec

kWh

Copyright©2016,Oracleand/oritsaffiliates.Allrightsreserved.|

• “Features” that help us estimate heat type: – Difference between winter gas usage and shoulder gas usage – Ratio between winter gas usage and shoulder gas usage – Difference between winter elec usage and shoulder elec usage – Ratio between winter elec usage and shoulder elec usage

Estimating heat type

0

5

10

Jan

Feb

Mar

A

pr

May

Ju

n Ju

l A

ug

Sep

O

ct

Nov

D

ec

Therms

0 10 20 30 40

Jan

Feb

Mar

A

pr

May

Ju

n Ju

l A

ug

Sep

O

ct

Nov

D

ec

kWh

Copyright©2016,Oracleand/oritsaffiliates.Allrightsreserved.|

Estimating heat type

Copyright©2016,Oracleand/oritsaffiliates.Allrightsreserved.|

Estimating heat type – suggestion from a client We got a couple complaints from XECO customers The client suggested some improvements…

“significant bump”

“relatively higher”

“significant increase”

Copyright©2016,Oracleand/oritsaffiliates.Allrightsreserved.|

• Want to estimate some value: – Does this household use GAS or ELECTRIC heat?

»  Have something we know about each household that might help us estimate

»  Know the answer for some instances

Standard machine learning setting

Copyright©2016,Oracleand/oritsaffiliates.Allrightsreserved.|

• Want to estimate some value: target variable »  Have something we know about each household that might

help us estimate: features • Know the answer for some instances: labeled training set

Standard machine learning setting

Copyright©2016,Oracleand/oritsaffiliates.Allrightsreserved.|

Training set

Copyright©2016,Oracleand/oritsaffiliates.Allrightsreserved.|

Goal: learn a function

Copyright©2016,Oracleand/oritsaffiliates.Allrightsreserved.|

Goal: learn a function

0 400 800

1,200 1,600 2,000

Jan

Feb

Mar

A

pr

May

Ju

ne

July

A

ug

Sep

O

ct

Nov

D

ec

Copyright©2016,Oracleand/oritsaffiliates.Allrightsreserved.|

Goal: learn a function input output

Copyright©2016,Oracleand/oritsaffiliates.Allrightsreserved.|

Goal: learn a function

Training consists of learning parameters or coefficients of our function

score = coeff1 * elec_diff + coeff2 * elec_ratio +

coeff3 * gas_diff + coeff4 * gas_ratio

function: if score > 0, then estimate is ELEC

otherwise estimate is GAS

positive

negative

Copyright©2016,Oracleand/oritsaffiliates.Allrightsreserved.|

Goal: learn a function score = coeff1 * elec_diff + coeff2 * elec_ratio +

coeff3 * gas_diff + coeff4 * gas_ratio

function: if score > 0, then estimate is ELEC

otherwise estimate is GAS

Coefficients quantify things like “relatively higher” Learned coefficients guaranteed to perform best on training set

Copyright©2016,Oracleand/oritsaffiliates.Allrightsreserved.|

Standard machine learning pipeline

Training Set Evaluation Set Real Life

train the function evaluate how well the function predicts

use the function on new data to get our

answers

Ja M M Jul

Se

No

coeff1: 1.38 coeff2: 0.25 coeff3: 3.59 coeff4: 2.84

Model accuracy: 86% Baseline accuracy: 72%

Copyright©2016,Oracleand/oritsaffiliates.Allrightsreserved.|

• Want to estimate some value: target variable – Can be category (ELEC/GAS) or number (e.g., kWh) – Category – classification; number – regression

»  Have something we know about each instance that might help us estimate: features

• Know the answer for some instances: labeled training set

Standard machine learning setting

The function you use doesn’t really matter The function we used earlier was logistic regression

Others include SVM, nearest neighbor, neural networks

Copyright©2016,Oracleand/oritsaffiliates.Allrightsreserved.|

• Target variable: – Will household X participate in program Y? (classification)

• Use case: – Program targeting

• Labeled training set: – We import past participation data from many clients

• Possible features: – Combination of demographic and behavioral data

Examples

Program Propensity

Copyright©2016,Oracleand/oritsaffiliates.Allrightsreserved.|

• Target variable: – Does household X have an EV? (classification)

• Use case: – Suggested rate changes, grid balancing

• Labeled training set: – EV rate code for some customers, other sources?

• Possible features: – Spikes in AMI data

Examples

Electric Vehicle Detection

Copyright©2016,Oracleand/oritsaffiliates.Allrightsreserved.|

• Target variable: – Will customer X call customer service? (classification)

• Use case: – Proactive messaging, call volume reduction

• Labeled training set: – Call center data?

• Possible features: – High bills, income level?

Examples

Customer Service Call Propensity

Copyright©2016,Oracleand/oritsaffiliates.Allrightsreserved.|

• Target variable: – Is household X a vacation home? (classification)

• Use case: – Better neighbor comparison, personalization

• Labeled training set: – Is this obtainable?

• Possible features: – Monthly or weekly usage patterns

Examples

Vacation home identification

Copyright©2016,Oracleand/oritsaffiliates.Allrightsreserved.|

• Target variable: – What is customer X’s business type? (classification)

• Use case: – Utilities want to know this; so does our SMB team

• Labeled training set: – SMB has 3rd party labels

• Possible features: – Business name, AMI usage

Examples

Business type classification

Copyright©2016,Oracleand/oritsaffiliates.Allrightsreserved.|

• Target variable: – What will usage be next week/month? (regression)

• Use case: – Bill forecast module; bill protect

• Labeled training set: – Past usage data

• Possible features: – Customer’s past usage, forecasted weather

Examples

Bill Forecasting

Copyright©2016,Oracleand/oritsaffiliates.Allrightsreserved.|

• Target variable: – What would usage be for given weather? (regression)

• Use case: – Weather normalization module

• Labeled training set: – Past usage/weather data

• Possible features: – Customer’s past usage, weather

Examples

Weather Normalization

Copyright©2016,Oracleand/oritsaffiliates.Allrightsreserved.|

• Everything we just saw was called “supervised learning” • What if we don’t have labeled data?

Unsupervised learning

Unsupervised Learning

Copyright©2016,Oracleand/oritsaffiliates.Allrightsreserved.|

• Unsupervised learning is looking for patterns in the data • Don’t know the right answer, and there is no “right answer” • E.g., clustering – how many clusters are there?

Unsupervised learning

Copyright©2016,Oracleand/oritsaffiliates.Allrightsreserved.|

• Unsupervised learning is looking for patterns in the data • Don’t know the right answer, and there is no “right answer” • E.g., clustering – how many clusters are there?

Unsupervised learning

Copyright©2016,Oracleand/oritsaffiliates.Allrightsreserved.|

• Unsupervised learning is looking for patterns in the data • Don’t know the right answer, and there is no “right answer” • E.g., clustering – how many clusters are there?

Unsupervised learning

Copyright©2016,Oracleand/oritsaffiliates.Allrightsreserved.|

• Unsupervised learning is looking for patterns in the data • Don’t know the right answer, and there is no “right answer” • E.g., clustering – how many clusters are there?

Unsupervised learning

Copyright©2016,Oracleand/oritsaffiliates.Allrightsreserved.|

Data Science at Opower Mission:

Generate valuable, actionable insights for utility clients and their customers and deliver them into our products

Methods: Employ machine learning techniques to develop customer models and behavioral understanding

Copyright©2016,Oracleand/oritsaffiliates.Allrightsreserved.|

Data Science workflow

Research •  Data exploration •  Accuracy testing •  Prototyping

Initial Rollout •  Professional Service •  Pilot

General Availability •  Productionalized as a service •  Available to all clients

Research •  Continued exploration •  Accuracy testing

Copyright©2016,Oracleand/oritsaffiliates.Allrightsreserved.|

MLAB Two goals:

(1) be a home for ongoing machine learning projects that are not part of any scrum (2) provide people an opportunity to be exposed to and learn about Machine Learning by participating in side projects

Ongoing Projects: Seasonal load curves, pool/washer/dryer detection, churn prediction, electric vehicle detection

Copyright©2016,Oracleand/oritsaffiliates.Allrightsreserved.|

PersonalizaConThroughLoadCurveAnalysis

Copyright©2016,Oracleand/oritsaffiliates.Allrightsreserved.|

Load Curves – All Customers

Copyright©2016,Oracleand/oritsaffiliates.Allrightsreserved.|

Load Curves – All Customers

Copyright©2016,Oracleand/oritsaffiliates.Allrightsreserved.|

Load Curves – All Customers

Copyright©2016,Oracleand/oritsaffiliates.Allrightsreserved.|

Load Curve Archetypes

Steady Eddies

Daytimers

Night Owls

0.00 4.00 8.00 12.00 16.00 20.00 24.00

3%

Prop

ortio

n of

usa

ge

in e

ach

hour

4%

5%

6%

Hour of the day 0.00 4.00 8.00 12.00 16.00 20.00 24.00

3%

Prop

ortio

n of

usa

ge

in e

ach

hour

4%

5%

6%

Hour of the day

0.00 4.00 8.00 12.00 16.00 20.00 24.00

3%

Prop

ortio

n of

usa

ge

in e

ach

hour

4%

5%

6%

Hour of the day

Evening Peakers

0.00 4.00 8.00 12.00 16.00 20.00 24.00

3%

Prop

ortio

n of

usa

ge

in e

ach

hour

4%

5%

6%

Hour of the day

Twin Peaks

0.00 4.00 8.00 12.00 16.00 20.00 24.00

3%

Prop

ortio

n of

usa

ge

in e

ach

hour

4%

5%

6%

Hour of the day

Copyright©2016,Oracleand/oritsaffiliates.Allrightsreserved.|

Segmentation

Copyright©2016,Oracleand/oritsaffiliates.Allrightsreserved.|

Targeted Messaging: Afternoon Peakers

This is an alert from UtilCo: Tomorrow, Wednesday, July 10th is a peak day.

From 2 PM to 7 PM join UtilCo customers by reducing your electric use. Simple ways to save on peak

days include postponing dishwashing and other large appliance use until the peak day is over. Thank you for helping

us save! To opt out of phone alerts, press 9.

Copyright©2016,Oracleand/oritsaffiliates.Allrightsreserved.|

Improved Personalization

Help drive acceptance of neighbor comparison

Copyright©2016,Oracleand/oritsaffiliates.Allrightsreserved.|

Improved Personalization

Recommendations tailored to profile type

Copyright©2016,Oracleand/oritsaffiliates.Allrightsreserved.|

ProgramPropensity

Copyright©2016,Oracleand/oritsaffiliates.Allrightsreserved.|

Target the right people with utility programs Target likely participants •  Some customers are more likely to

participate in any program

Target specific customers for certain programs •  Different types of customers are better

fitted for different utility programs, indicated by their propensity

•  Target low propensity customers for simple programs, and high propensity customers for more involved customers

High Propensity Program

Low Propensity Program

Copyright©2016,Oracleand/oritsaffiliates.Allrightsreserved.|

Underneath the hood

Load shape

$

Monthly usage

Web behavior

Income

Home data

PredicCvemodel

•  LiWparCcipaCon~20%•  DecreasemarkeCngspendthroughincreasingrelevance

Copyright©2016,Oracleand/oritsaffiliates.Allrightsreserved.|

EnergyDisaggregaConandSetpointEsCmaCon

Cooling

32%

Copyright©2016,Oracleand/oritsaffiliates.Allrightsreserved.|

Jan Apr Jul Oct Jan Apr Jul Oct

Baseload

Heating Cooling

Energy Disaggregation

Copyright©2016,Oracleand/oritsaffiliates.Allrightsreserved.|

Disaggregation at Opower

Copyright©2016,Oracleand/oritsaffiliates.Allrightsreserved.|

Beyond Heating/Cooling Disaggregation Learn more about individual homes using just energy usage data (e.g., AMI, bills)

Copyright©2016,Oracleand/oritsaffiliates.Allrightsreserved.|

Setpoint Detection

base load

cooling setpoint

one household

one hour

Copyright©2016,Oracleand/oritsaffiliates.Allrightsreserved.|

Setpoint Detection

cooling setpoint - 88°

Copyright©2016,Oracleand/oritsaffiliates.Allrightsreserved.|

Setpoint Detection

cooling setpoint - 76°

Copyright©2016,Oracleand/oritsaffiliates.Allrightsreserved.|

Setpoint Detection

cooling setpoint - 64°

Copyright©2016,Oracleand/oritsaffiliates.Allrightsreserved.|

Setpoint Detection

cooling setpoint - 79°

heating setpoint - 62°

Copyright©2016,Oracleand/oritsaffiliates.Allrightsreserved.|

Setpoint Detection – Hourly Analysis

Copyright©2016,Oracleand/oritsaffiliates.Allrightsreserved.|

Setpoint Detection – Hourly Analysis

For any given temperature and hour of the day, what percentage of total usage is

due to cooling?

Copyright©2016,Oracleand/oritsaffiliates.Allrightsreserved.|

Setpoint Detection – hourly analysis

Copyright©2016,Oracleand/oritsaffiliates.Allrightsreserved.|

Accurate Disaggregation

Copyright©2016,Oracleand/oritsaffiliates.Allrightsreserved.|

Tip Targeting

Copyright©2016,Oracleand/oritsaffiliates.Allrightsreserved.|

Household Targeting For DR Event

Setpoint: 74° Event savings: 3 kWh DR: MAYBE

Setpoint: 79° Event savings: 0.5 kWh DR: NO

Setpoint: 68° Event savings: 5.5 kWh DR: YES

Copyright©2016,Oracleand/oritsaffiliates.Allrightsreserved.|

Bill Forecasting

Copyright©2016,Oracleand/oritsaffiliates.Allrightsreserved.|

• Identify machine learning problems – If you ever say to yourself, I wish I knew X about a customer, maybe we can! – Come talk to the Data Science team

• Is your problem supervised or unsupervised? – Do we have data with labeled target variables?

• We’re about to get more data – What can we do with it? – What new problems are there? – What old problems can better be solved with new data?

Conclusions

Copyright©2016,Oracleand/oritsaffiliates.Allrightsreserved.|

71

Thanks!