Federated machine learning for creating power load profiles

The project explores how federated machine learning can create dynamic load type profiles for a digital forecasting tool for future grid capacity.

Photo: Tomas Ärlemo, Svenska kraftnät

AI Sweden has been investigating and developing knowledge around federated machine learning on sensitive datasets in health, vehicle data, and financial transaction data for several years. This technology enables training on datasets across multiple parties without data sharing.

Svenska Kraftnät has a government mandate to plan for increased electricity use and needs to develop long-term planning. Accurate forecasts are required to visualize where electricity production, flexibility resources, and electricity usage should be connected.

This project investigates whether federated machine learning on electricity grid data can create dynamic load type profiles for a digital forecasting tool for future grid capacity. The project will also investigate the value of machine learning compared to classical methods and identify potential institutional barriers.

Challenges

The main challenge is how electricity grid companies can collaborate to create a common ML model for load type profiles without violating legislation. The project will also create a mockup of a user-friendly tool for visualizing available grid capacity.

Project purpose

The purpose is to investigate if and how federated machine learning on electricity grid data can contribute to the creation of dynamic load type profiles for a digital forecasting tool for future available grid capacity.

Expected outcomes

A functioning federated “learning infrastructure” of ML models
Selection of format and structure for necessary electricity grid data and metadata
A prototype of an ML model that generates synthetic load type profiles
A “mockup” of a capacity map and user interface

What is Federated machine learning?

Federated machine learning is a technique where multiple stakeholders collaboratively train a machine learning model without sharing their raw data. Instead, each device or silo trains the model locally and only shares updates (like gradients), keeping the data private and secure.