Continuous Deep Learning for Smarter Energy Igor Gotlibovych - - PowerPoint PPT Presentation

Continuous Deep Learning for Smarter Energy

Igor Gotlibovych Javier Asensio-Cubero Jaye Cribb Constantin Octopus

The Energy is Getting Smarter

Traditional grid:

• predictable generation
• static consumption
• fixed tariffs
• timescale of years/months

Smart grid:

• distributed renewable generation
• dynamic consumption behaviour
• smart meters
• agile tariffs
• API integration
• timescale of hours/minutes

2

A Brief History of Octopus Energy & AI

• founded in 2015
• Apr 2017: 100% renewable sources
• Feb 2018: UK’s first Agile tariff
• Jul 2018: +100k IRESA customers
• Nov 2018: acquired ML team from Usio
• Jan 2019: 500k customers
• ~600k customers; ~50k smart meters
• 20+ developers
• 6+ Data Scientists/ML Engineers

Machine Learning at Octopus:

• Time Series Analysis & Forecasting:

○ household consumption ○ response to agile tariffs ○ commercial sites ○ PV generation ○ Customer communications: calls/emails

• NLP
• Control theory/RL

○ Optimizing PV/battery/EV utilization

• Growing ETL/Deep Learning infrastructure

3

Smart tariffs and Behavioural Response

Agile tariff response: early trial Agile tariff response analysis: regression + SHAP

4

Time Series Forecasting with Deep Learning

“What will the half-hourly consumption be for customer #500001 at 2019-02-29 12:00?” features X prediction y 500001 2019 2 29 12 2.3

• Why not iterative auto-regressive model?

○ Deep learning is powerful enough to turn this: f(f(f(f(...f(X)...)))) [n times] into this: g(X, n)

5

6

• Keep continous features continuous
• Better still, use periodic encoding

Feature Engineering - Datetime

Jan-31 12:00 0.084 0.5

7

Feature Engineering - IDs

• You have a categorical feature
• You decide to use one-hot encoding
• Now you have 500k+ features!

500001 ... 1

8

What we really want to know

#500001 What time do they come home? Do they have electric heating? Do they work weekends? ... m features

9

Embeddings: the MovieLens Rating Example

10

Putting it Together

11

What do Embeddings Learn?

Some things we can’t predict: CTO decides to eat out Portfolio Individual

12

Model Life-cycle & Continuous Training

time

13

Continuous deployment, training, and usage.

• Like cron on steroids: Flexible way to schedule tasks, rerun them, adds

visibility…

○ Written in python. ○ Rather a monolithic approach: code base sharing for the tasks.

Heterogeneous architecture

Final approach

• Dockerized workers

○ Decoupling airflow of dependencies and scheduling. ○ Decoupling airflow from credentials management.

• AWS autoscaling

○ Turning on and off GPUs on demand.

• More maintainable codebase.

○ Based on command line scripts. ○ Highly testable and easy to isolate. ○ Airflow dags relegated to a glorified configuration file.

Lessons Learned

• Avoid iterative forecasting - use features instead
• Two ways to learn latent features

○ Embeddings: learn features during backprop ○ Encoders: trained to extract features in forward pass (not covered)

• Encapsulate feature generation inside your model

○ simpler versioning and deployment ○ simplifies replacing hand-engineered features with end-to-end ML models

• Build reusable ML code

○ Grow your toolkit ○ Listen to ML Engineers when they tell you to write tests

• Automate every part of the model lifecycle (Airflow, Kubernetes, ASGs, …)

17

Thanks!

igor [at] octopus.energy https://ig248.gitlab.io/ @igorgotlibovych javier.asensio [at] octopus.energy Should Data Scientists write Unit Tests? Switch to Octopus Energy here: https://share.octopus.energy/mist-rain-538 Switch to Octopus Energy here: https://share.octopus.energy/sand-fawn-149 always! maybe?

18