Future-proof energy systems with ML: tackling data drift & low-data forecasting

DETAILS

It's almost exactly one year ago; time for another co-organized meetup with Dexter Energy on November 28 at 18:00 CET at their office in Amsterdam - Joan Muyskenweg 22, 1096 CJ Amsterdam. On this meetup you'll learn more about how to use Bayesian modeling for sparse data coverage, and additionally different methodologies and challenges of detecting data drift will be discussed.

Get energized by practical data science use cases in the space of forecasting and data drift, knowledge sharing and of course networking; see you all soon!

SCHEDULE

18:00 - 19:00: Walk-in with drinks and bites & welcome!
19:00 - 19:45: Using Bayesian modeling to Overcome Low Data Coverage by Aletta Heemskerk
19:45 - 20:00: Break
20:00 - 20:45: Adapting to change: dealing with data drift in energy waste prevention by Tatjana Puskarov
20:45 - 22:00: Networking and drinks

TALKS

[Talk 1]: Using Bayesian modeling to Overcome Low Data Coverage by Aletta Heemskerk

A key part of forecasting wind power generation is fitting the wind speed-energy curve. There are various ways of estimating this relation. Common options are a parametric model, basically a good baseline, or using Machine Learning.

However, some of these methods can be prone to outliers or struggle when limited samples are available. Alternatively, one can use Bayesian modeling. This approach yields easy-to-interpret results and allows one to handle sparse data coverage.

In this talk, we will start with the foundation of Bayesian statistics and introduce the problem at hand. We will cover how to apply HMC NUTS using NumPyro including the specification, fitting, and validation using the business case of forecasting wind energy in low data regimes.

[Talk 2]: Adapting to change: dealing with data drift in energy waste prevention by Tatjana Puskarov

At Sensorfact we aim to reduce the human environmental footprint by eliminating all waste in industry, a large part of which comes from idling machines. We have set up a machine learning system to do this based on real time power consumption data. The machine-level models underlying this system are trained on a reference period and continuously predict in near real time. However, as machines wear out or get serviced, or as our customers make changes in their production, the data drifts - leading to a diminishing performance of our ML system and potentially unnoticed preventable standby consumption. We have therefore implemented data drift detection in the system lifecycle - with data drift alerts, we can retrain our models in time to ensure their quality. In this talk, we will discuss the methodologies and challenges of detecting drift and the solution that was selected for our use case.

NOTE:
Our team will be ready to assist with check-in starting at 18:00, so please arrive a bit early if possible.

Since we can only host a limited amount of attendees for this meetup event, if you cannot join the event, update your status to "not going" so you can give your spot to the folks on the waitlist. Thanks for your help!

DIRECTIONS

Address:
Dexter Energy Amsterdam Office
Joan Muyskenweg 22, 1096 CJ Amsterdam

By Bike:
If you’re already in Amsterdam or nearby, biking is a convenient option. Use Google Maps or a similar app for a bike-friendly route. Most paths will guide you to Amstelkwartier west.

By Public transport:
Take a metro to Amsterdam Overamstel Station. The Dexter Energy office is located next to the Overamstel metro station; it's just a 250m walk.