Trust-ML: An Optimization-based Platform for Building Trust in Machine Learning Models used for Power Systems

Projektbeschreibung

Auf dem Weg zu vertrauenswürdigem maschinellem Lernen

Mithilfe des maschinellen Lernens ist es möglich, die tiefgreifende Dekarbonisierung des Energiesektors voranzutreiben. Dank seiner Fähigkeit, in komplexen Umgebungen zu lernen und Lösungen zu finden, kann maschinelles Lernen die Energiesysteme grundlegend verändern. Die sich abzeichnenden EU-Verifizierungsstandards verlangen jedoch, dass sämtliches in sicherheitskritischen Anwendungen eingesetztes maschinelles Lernen sowie Verstärkungslernen nachweislich vertrauenswürdig ist. Das EU-finanzierte Projekt TRUST-ML zielt auf die Entwicklung eines einheitlichen Rahmens für die Bewertung der quantitativen Vertrauenswürdigkeit von Modellen neuronaler Netze ab, die üblicherweise in Energiesystemen eingesetzt werden. TRUST-ML wird einen neuartigen, konvexen Optimierungsansatz anwenden, um die Vertrauenswürdigkeit von maschinellem Lernen in Bezug auf Leistung, Robustheit und Interpretierbarkeit zu bewerten. Es ist zudem auf die neuen Anforderungen der aktuellen Stromnetze ausgerichtet.

Ziel

Deep decarbonization of the energy sector will require massive penetration of stochastic renewable energy resources and an enormous amount of grid asset coordination; this represents a challenging paradigm for power system operators. With its ability to learn in complex environments and provide predictive solutions on fast timescales, machine learning (ML) is posed to help overcome these challenges and dramatically transform power systems in coming decades. Emerging EU verification standards, however, will require that all ML and Reinforcement Learning (RL) used in safety critical applications be demonstrably trustworthy. In this project, we develop a unified framework, known as Trust-ML, for assessing the quantitative trustworthiness of the neural network models commonly used in power systems. Trust-ML uses a novel, convex optimization approach to assess ML trustworthiness across three key dimensions: performance, robustness, and interpretability. The approach is engineered to be scalable, and by design, it generates exact verification guarantees. Furthermore, Trust-ML is designed to meet the emerging needs of actual power systems. In particular, it can verify the performance of multi-agent RL systems in rigorous ways, and its relaxed counterpart can offer tractable, worst-case performance guarantees in the context of online learning. The resulting verification tools will be published as open-source software packages and shared widely with researchers and industry. This project will advance state-of-the-art methods across several interdisciplinary fields, it will help remove the barriers associated with machine learning deployment in power systems, and its outcomes will help push European power grids into competitive spaces. Coming from MIT with advanced training in power systems, the project PI, Samuel Chevalier, is characteristically well-suited to build Trust-ML, and his team of advisors represents a mixture of experts across power, optimization, and learning.