GridAssist: AI-based congestion mangement and grid-restoration

GridAssist – Assistenzsysteme für eine optimierte automatisierte Systemführung in Verteilnetzen

This project is funded by the Bundesministerium für Wirtschaft und Energie (BMWE).

On this page:

• Abstract
• Research Focus
• Use Cases
• Validation and Field Test
• Publications
• Project Details

Abstract

A central goal of GridAssist is the development of machine-learning-based methods for bottleneck and fault management. These methods build on an integrated knowledge base that combines heterogeneous power-grid data sources, including topology information, operational measurements, and contextual data. Different AI approaches are investigated, including supervised and unsupervised learning, as well as hybrid strategies with generic pre-training and later adaptation to the real network situation.

To ensure safety, robustness, and physical plausibility, the project further investigates Known-Operator Learning and Physics-Informed Neural Networks (PINNs). By incorporating prior physical knowledge about the power grid directly into the learning process, these approaches restrict the solution space to physically meaningful actions, reduce the amount of required training data, and improve the reliability of the generated recommendations.

For congestion management, the assistance system is triggered by tolerance-band violations or anomaly detection, both in current and predictive settings. It then proposes suitable remedial actions such as topology changes and the targeted use of flexibilities in the infeed and outfeed areas, while considering operational and economic constraints such as voltage limits, short-circuit strength, neutral point treatment, (n-1) security, and cost functions.

For fault management and restoration, GridAssist supports the identification of fault locations, the isolation of affected network areas, and the restoration of supply to unaffected sections of the grid. The project builds on existing field measurements and indicators, while also exploring additional measurements and methods to improve the quality and robustness of restoration strategies.

The developed methods are validated in a hybrid field-test environment. Realistic disturbance and congestion scenarios are reconstructed from field data and simulated in real time using an OPAL-RT environment. Real-time grid snapshots in CGMES format are processed and connected to the assistance systems via standard control-center protocols, enabling safe and vendor-independent evaluation under realistic operating conditions.

Research Focus

• Automated congestion management based on anomaly detection, topology reconfiguration, and flexibility activation
• Fault management and service restoration with AI-supported localization, switching, and restoration strategies
• Knowledge-based integration of heterogeneous power-grid data using graph-based data models and databases
• Safe and physically meaningful AI through Known-Operator Learning and Physics-Informed Neural Networks
• Real-time validation using realistic field data, CGMES snapshots, and OPAL-RT-based simulation

Use Cases

Automated Bottleneck Management:
The system detects congestion situations caused by tolerance-band violations or anomalous network states and proposes suitable countermeasures such as topology changes or the use of flexibilities. These recommendations must remain technically feasible and economically meaningful under realistic operational constraints.

Fault Management and Grid Restoration:
The system supports fault localization, isolation of faulted network sections, and restoration of service to unaffected areas. It is designed to assist operators in increasingly complex network situations and to improve resilience and continuity of supply.

Knowledge Database for Heterogeneous Grid Data:
GridAssist develops a graph-based knowledge database that integrates topology, SCADA, metering, and contextual data. This enables efficient querying of relationships between assets, measurements, and operational states and provides the data backbone for the AI-based assistance functions.

Validation and Field Test

The assistance systems developed in GridAssist are evaluated in a hybrid field-test environment. Since critical interventions cannot be tested directly on live infrastructure to the required extent, realistic scenarios are reconstructed from field measurements and simulated in real time.

For this purpose, real-time system snapshots in CGMES format are provided from the control-center environment over an extended period, processed by a dedicated converter, and fed into an OPAL-RT simulator. The real-time simulator communicates with the assistance systems via standard control-center protocols, enabling secure and realistic evaluation of the proposed methods. Experienced operators are involved in the validation in order to assess the practical applicability of the generated recommendations.

This setup ensures that the developed solutions remain vendor-independent and transferable to a broad range of grid operators.

Publications

Project Details

• Project acronym: GridAssist
• Project title: Assistenzsysteme für eine optimierte automatisierte Systemführung in Verteilnetzen
• Funding: Bundesministerium für Wirtschaft und Energie (BMWE)
• Project duration: 01.12.2024 – 30.11.2027
• CRIS project ID: 351939872

2024

• Teilvorhaben: KI gestütztes Engpassmanagement und Netzwiederaufbau

  
  (Third Party Funds Group – Sub project)
  Overall project: GridAssist - Assistenzsysteme für eine optimierte automatisierte Systemführung in Verteilnetzen
  Project leader: Siming Bayer, Gert Mehlmann, Johann Jäger, Matthias Luther
  Term: December 1, 2024 - November 30, 2027
  Acronym: GridAssist
  Funding source: Bundesministerium für Wirtschaft und Energie (BMWE)
  
  Abstract

  TP3: Automated Optimal System Management in MS/HS Networks & Network Restoration

  AP 3.1: Automated Bottleneck and Fault Management Based on Machine Learning Methods

  AP 3.2: Stable Grid and Supply Restoration

  
  

  TP5: Competence Team for Interfaces and Databases

  AP 5.1: Creation of a Knowledge Database for Heterogeneous Power Grid Data Sources

  Automated Grid Management and Restoration through AI and Data Integration

  The increasing complexity and volatility of modern power grids necessitate intelligent, automated systems for optimal operation and rapid restoration. In TP3, we develop AI-driven solutions for automated bottleneck and fault management (AP 3.1) as well as stable and systematic grid restoration (AP 3.2) in medium- and high-voltage (MS/HS) networks. These systems leverage machine learning techniques—ranging from supervised and unsupervised learning to reinforcement learning—with a hybrid training approach incorporating domain knowledge via Physics-Informed Neural Networks (PINNs) and Known-Operator Learning. This ensures reliable, explainable decision-making in real-time operations and minimizes data requirements in safety-critical environments.

  In bottleneck management, predictive and event-triggered AI models autonomously suggest topology adjustments and flexibility deployment, factoring in a wide range of grid constraints and operational targets. Fault management employs multi-source data analysis for real-time fault detection, location, isolation, and resupply, utilizing both existing sensors and proposed novel measurements. For restoration, a dedicated assistance system is developed using a two-stage training process: supervised learning of standardized grid restoration procedures followed by reinforcement learning in a high-fidelity simulation environment. A single-agent model competes across simulation instances to evolve optimized strategies for power system recovery.

  
  

  To support these applications, TP5 (AP 5.1) establishes a unified graph-based knowledge database for integrating heterogeneous data sources—from GIS, SCADA, and metering systems—into a cohesive representation. This structure enables complex system queries and machine-readable context for AI models, forming the backbone of advanced operational analytics.

  The entire system architecture and algorithms are tested in AP 6.3 through a hybrid field test utilizing real-time simulation with OPAL-RT and live grid snapshots from Lechwerke’s control center. Assistance systems interface with the simulation via standard protocols, ensuring vendor-independent implementation and practical transferability. Experienced grid operators validate the system in realistic operational conditions, ensuring alignment with the requirements of future grid operations and energy transition goals.

  →More information