Saving Energy in Homes
with a Unified Approach to
Data and AI
Stephen Galsworthy, Chief Data & Product Strategy Officer, Quby
www.linkedin.com/in/galsworthy/
Erni Durdevic, Tech Lead – Data Science & Engineering, Quby
www.linkedin.com/in/ernidurdevic/
Our vision:
A world where living without
wasting natural resources is easy
Source: Eurostat
44%
Households use 44% of all natural gas
and 27% of all electricity in the EU
27%
But how much is just wasted?
We are not here to tell
people how much to
consume …
… but to make sure all their
use is intended.
400,000 connected homes across Europe
Data Technology Stack: overview
IoT and Customer data Quby platform Personalised services
Cloud data storage AI control centre
• Advice & Insights
• Smart thermostat
control
• Home monitoring
Waste Checker
• We have found that over 40% of household appliances are inefficient or
being used inefficiently
“It looks like you often wash at high
temperatures. Your dishwasher needs
more energy to heat the water.
Washing at 50 degrees, or ECO mode, will
get your utensils just as clean, go ahead
and try it.”
Waste Checker offers personalised advice on
inefficient appliances and behaviours
Source: EU
Central heating
system
Water sensorElectricity sensorGas sensor
IoT Data Models Personalised
services
Terabytes of IoT data daily
Our data-driven services
Challenges
3. Over 1 million unsupervised
models trained daily
4. Struggling to track model
performance and
training/test datasets for a
wide range of algorithms
IoT Data Models Personalised
services
Patented algorithms
Our data-driven services
DishwasherWashing
machine
Washing
machine
DryerDryer
Patented algorithms
Personalised advice
on how to avoid
wasting energy
IoT Data Models Personalised
services
Our data-driven services
Batch processing
Our unified data analytics setup
IoT stream
Click data
Exploratory Data
Science Notebooks
Real time services
Live dashboards
SQL
Services API
SQL / NoSQL
AWS S3
Batch
Streaming
R & D
Model deployment
Model registry
Batch processing
Our unified data analytics setup
IoT stream
Click data
SQL
Services API
SQL / NoSQL
AWS S3
>1 TB per
day > 3 PB
Batch
>1 Million
unsupervised
models trained
daily
Batch processing
Unit tests
Electricity
Water
Presence
detection
Schedule
advice
…
Batch to Streaming
Batch
…
…
Streaming
• The same transformation will work for
Select
Filter
WithColumn
GroupBy
Join
…
UDF
Time-based window
Non time-based window functions (lead, lag,
…)
Batch processing
Our unified data analytics setup
IoT stream
Click data
Exploratory Data
Science Notebooks
SQL
Services API
SQL / NoSQL
AWS S3
https://github.com/quby-io/databricks-workflow
Demo – Who is it for?
Data Science teams that want to
get some IoT data pipelines in
production with Databricks.
https://github.com/quby-io/databricks-workflow
Demo – Why?
Help you to figure out a simple way to answer to the questions
• How do I set up a production pipeline in Databricks?
• How do I run unit tests on Spark transformations?
• How do I run integration tests on Notebook workflows?
• How do I rapidly prototype data transformations on real data?
• How do I manage multiple environments and configurations?
Why is it so special?
• Because it gives an insider view on a well-rounded way of working.
• It has multiple best practices pre-baked into a working example.
https://github.com/quby-io/databricks-workflow
Demo Databricks workflow
staging
raw
staging _features
electricity_power
dev_erni_features
electricity_power
power_features
initialize create_features
production
raw
production_features
electricity_powerinitialize create_features
integration_test
raw
integration_test_features
electricity_powerinitialize create_features
https://github.com/quby-io/databricks-workflow
Batch processing
Our unified data analytics setup
IoT stream
Click data
Exploratory Data
Science Notebooks
Real time services
Live dashboards
SQL
Services API
SQL / NoSQL
AWS S3
>1 TB per
day > 3 PB
>500 models
logged
Batch
Streaming
R & D
Model deployment
Model registry
>1 Million
unsupervised
models trained
daily
>15
deployments
tracked
Three crucial aspects can be simplified by applying
MLflow and model registry
1
Choosing the right model
2
Finding the latest model
3
Deploying the model to production
The source code of a model is not just the
code, but the combination of code,
configuration and data.
Choosing the right model before
?
Choosing the right model with MLflow
Finding the latest model before
?
”Hey Erni, where’s
the latest model?”
Finding the latest model with MLflow model registry
Model deployment before
Model deployment with MLflow model registry
Unified data analytics is powering Quby’s
transformation into a AI-first company
• Less time on spent on managing data infrastructure
• Collaboration and re-use of code and models across all
product teams
• Uniform way of working, including logging metrics and
tracking performance in MLflow
• Faster development cycles with new features delivered to
end users every 2 weeks
End result: Saving energy in homes across Europe
• Waste Checker, in the last 12 months:
• 87 million inefficient appliance cycles identified
• 67,000,000 kWh of wastage identified and targeted
• Thermostat Program Advice:
• Users automatically alerted when using gas whilst not at home
• 87,500 m3 of gas saved last winter
• Step-by-step we are enabling the transition to a sustainable energy system
Saving Energy in Homes with a Unified Approach to Data and AI

Saving Energy in Homes with a Unified Approach to Data and AI

  • 1.
    Saving Energy inHomes with a Unified Approach to Data and AI Stephen Galsworthy, Chief Data & Product Strategy Officer, Quby www.linkedin.com/in/galsworthy/ Erni Durdevic, Tech Lead – Data Science & Engineering, Quby www.linkedin.com/in/ernidurdevic/
  • 2.
    Our vision: A worldwhere living without wasting natural resources is easy
  • 3.
    Source: Eurostat 44% Households use44% of all natural gas and 27% of all electricity in the EU 27% But how much is just wasted?
  • 4.
    We are nothere to tell people how much to consume … … but to make sure all their use is intended.
  • 5.
  • 6.
    Data Technology Stack:overview IoT and Customer data Quby platform Personalised services Cloud data storage AI control centre • Advice & Insights • Smart thermostat control • Home monitoring
  • 7.
    Waste Checker • Wehave found that over 40% of household appliances are inefficient or being used inefficiently “It looks like you often wash at high temperatures. Your dishwasher needs more energy to heat the water. Washing at 50 degrees, or ECO mode, will get your utensils just as clean, go ahead and try it.” Waste Checker offers personalised advice on inefficient appliances and behaviours Source: EU
  • 8.
    Central heating system Water sensorElectricitysensorGas sensor IoT Data Models Personalised services Terabytes of IoT data daily Our data-driven services
  • 9.
    Challenges 3. Over 1million unsupervised models trained daily 4. Struggling to track model performance and training/test datasets for a wide range of algorithms IoT Data Models Personalised services Patented algorithms Our data-driven services DishwasherWashing machine Washing machine DryerDryer Patented algorithms
  • 10.
    Personalised advice on howto avoid wasting energy IoT Data Models Personalised services Our data-driven services
  • 11.
    Batch processing Our unifieddata analytics setup IoT stream Click data Exploratory Data Science Notebooks Real time services Live dashboards SQL Services API SQL / NoSQL AWS S3 Batch Streaming R & D Model deployment Model registry
  • 12.
    Batch processing Our unifieddata analytics setup IoT stream Click data SQL Services API SQL / NoSQL AWS S3 >1 TB per day > 3 PB Batch >1 Million unsupervised models trained daily
  • 13.
  • 14.
    Batch to Streaming Batch … … Streaming •The same transformation will work for Select Filter WithColumn GroupBy Join … UDF Time-based window Non time-based window functions (lead, lag, …)
  • 15.
    Batch processing Our unifieddata analytics setup IoT stream Click data Exploratory Data Science Notebooks SQL Services API SQL / NoSQL AWS S3 https://github.com/quby-io/databricks-workflow
  • 16.
    Demo – Whois it for? Data Science teams that want to get some IoT data pipelines in production with Databricks. https://github.com/quby-io/databricks-workflow
  • 17.
    Demo – Why? Helpyou to figure out a simple way to answer to the questions • How do I set up a production pipeline in Databricks? • How do I run unit tests on Spark transformations? • How do I run integration tests on Notebook workflows? • How do I rapidly prototype data transformations on real data? • How do I manage multiple environments and configurations? Why is it so special? • Because it gives an insider view on a well-rounded way of working. • It has multiple best practices pre-baked into a working example. https://github.com/quby-io/databricks-workflow
  • 18.
    Demo Databricks workflow staging raw staging_features electricity_power dev_erni_features electricity_power power_features initialize create_features production raw production_features electricity_powerinitialize create_features integration_test raw integration_test_features electricity_powerinitialize create_features https://github.com/quby-io/databricks-workflow
  • 19.
    Batch processing Our unifieddata analytics setup IoT stream Click data Exploratory Data Science Notebooks Real time services Live dashboards SQL Services API SQL / NoSQL AWS S3 >1 TB per day > 3 PB >500 models logged Batch Streaming R & D Model deployment Model registry >1 Million unsupervised models trained daily >15 deployments tracked
  • 20.
    Three crucial aspectscan be simplified by applying MLflow and model registry 1 Choosing the right model 2 Finding the latest model 3 Deploying the model to production
  • 21.
    The source codeof a model is not just the code, but the combination of code, configuration and data.
  • 22.
    Choosing the rightmodel before ?
  • 23.
    Choosing the rightmodel with MLflow
  • 24.
    Finding the latestmodel before ? ”Hey Erni, where’s the latest model?”
  • 25.
    Finding the latestmodel with MLflow model registry
  • 26.
  • 27.
    Model deployment withMLflow model registry
  • 28.
    Unified data analyticsis powering Quby’s transformation into a AI-first company • Less time on spent on managing data infrastructure • Collaboration and re-use of code and models across all product teams • Uniform way of working, including logging metrics and tracking performance in MLflow • Faster development cycles with new features delivered to end users every 2 weeks
  • 29.
    End result: Savingenergy in homes across Europe • Waste Checker, in the last 12 months: • 87 million inefficient appliance cycles identified • 67,000,000 kWh of wastage identified and targeted • Thermostat Program Advice: • Users automatically alerted when using gas whilst not at home • 87,500 m3 of gas saved last winter • Step-by-step we are enabling the transition to a sustainable energy system