Synthetic Topology Generation for Generalizable Machine Learning in Low Voltage Distribution Grids

SDEWES 2025

Markus de Koster
markus.de_koster@th-koeln.de

TH Köln - Cologne University of Applied Sciences

Agenda

workflow preview

Background

Methodology

Dataset

Conventional Power Grids

  • No generators in LV section
  • Loads are predictable
  • Minimal monitoring required

Transition to Renewable Energy

  • Distributed Generation
  • More significant loads in LV section
  • Additional monitoring required

Use of Artificial Intelligence

Money icon Monitoring everything is too expensive

Neural network icon Artificial Intelligence (AI) can infer unmonitored states

History Line Icon But: AI requires training

Reptitive Training

  • Grids vary in complexity and structure
  • Requires training of individual models per grid
  • Not sustainable

Generalizable Machine Learning

  • Train a single model on many different grids
  • Model learns general patterns
  • Can be applied to new grids without retraining
  • But: Requires many different training grids

This work provides a framework that generates synthetic power grid graphs for training of robust ML models

Why synthetic grids?

Secure note icon Real grid data is sensitive

database icon Few public datasets

structure icon Grids of different structure needed

scale icon Equal distribution of different structures benefits training

Graph structural properties

What do distribution grids typically look like?

  • Long radial feeders
  • Multiple loads
  • Few interconnections (open ring nets)

Source: Dickert et al., “Benchmark low voltage distribution networks based on cluster analysis of actual grid properties” 2013

Structural properties

Sources: Hu et al., “Open Graph Benchmark: Datasets for Machine Learning on Graphs” 2021

Fraunhofer IEE, “Power System Test Cases - pandapower 2.14.9” accessed 07/2024

Methodology

Core: Random Tree Model

with probabilities for

structure icon Branching

electric car icon Loads

solar panel icon Generators

Markov Chain

Markov Chain

  • Reflects real-world neighborhood structures
  • Ensures probabilities are met in the long run

Creating the graph structure

Assigning electrical parameters

electric car icon Add loads based on

  • area (urban, suburban, rural)
  • structure (residential, commercial, industrial)

solar panel icon Add generators

power pole icon Size lines and transformers based on expected load

Validation summary

All grids are verified through a power-flow simulation

Dataset summary
Different topologies 9,976
Load-flow converged 7,101
– rural 2,848
– suburban 2,390
– urban 1,863

Evaluation

Evaluation

Evaluation

Summary

history line icon Efficient method for synthetic grid generation

structure icon Generates realistic yet diverse grids

neural network icon Future work: train a ML model on generated grids

Interested in the dataset?

QR Code of contact info

Contact:

markus.de_koster@th-koeln.de

Or scan the QR code to get in touch