Universidad Pontificia Comillas. Madrid (España)
26 de junio de 2026
Resumen:
The transition to renewable energy and the simultaneous aging of electrical grid assets present a large-scale challenge to managing electrical power system assets. The traditional approach to asset management is based on subjective expert opinion and time-based preventive maintenance; however, these methods are becoming increasingly less effective due to the increased complexity and heterogeneity of modern power networks, as well as the growing size of these networks. A key objective of this dissertation is to develop an
intelligent environment for asset management utilizing machine learning methodologies that provides scalable, objective, and easily interpretable solutions.
This research was organized into six major research questions, which were answered through three peer-reviewed articles that are central to this dissertation. This study first identified 8 critical gaps in applying Machine Learning to Asset Management within the literature of over 150 references, which included the lack of integrated frameworks across multiple dimensions of asset assessment, limited discussion of data quality impacts on ML applications, over-reliance upon subjective expert judgement, and little access to open source resources for implementation of these technologies. Next, this study developed and tested an open source toolbox that implements a multi-dimensional evaluation framework of Life Assessment, Health Condition, Maintenance Strategy and Economic Impact assessments across five different types of assets (power transformers, Load Tap Changers, Circuit Breakers, Overhead Transmission Lines and Underground Cables), that are based upon five different dimensions of assessment. The toolbox consists of three modules: Module I utilizes unsupervised learning (K-Means Clustering and Self-Organizing Maps) to characterize assets objectively and Module II calculates standardized health indicators and aggregates them into a single indicator of total asset condition for ranking purposes. Module III utilizes reinforcement learning (Q-Learning) to develop the optimal maintenance strategy for each asset. Finally, additional advancements were made to address practical limitations associated with deploying ML-based solutions into operational environments, including the
systematic testing of methodologies for imputing missing data for handling missing data, the development of data-driven methodologies for weighting features as an alternative to relying upon subjective expert judgement, and the use of meta-optimization frameworks for solving multi-objective problems related to planning maintenance strategies.
Extensive validation across multiple case studies demonstrates the effectiveness and generalizability of the developed methodologies. This study will contribute to the development of sustainable and resilient electrical distribution systems, which will extend the operational life of critical assets, improve the efficient use of resources, reduce environmental degradation caused by early replacement of equipment, and support the development of reliable grids that can be affected by an increased amount of renewable energy. The
open source nature of the toolbox is intended to foster innovation among communities (community-driven) and help bridge the existing knowledge gap between academia and industry with respect to technology transfer. The future directions for this research may include incorporating additional asset classes into the framework, incorporating symbolic data and/or deep learning algorithms to improve the performance of the framework, using hierarchical/multi-agent reinforcement learning to optimize the entire fleet at once, and conducting long-term longitudinal testing to evaluate the ability of the framework to predict failure events.
Resumen divulgativo:
Esta investigación utiliza inteligencia artificial para ayudar a las empresas eléctricas a supervisar el estado de sus equipos, detectar problemas con antelación y planificar el mantenimiento de forma más eficiente, reduciendo costes y mejorando la fiabilidad de la red.
Descriptores: Inteligencia Artificial, Transmisión y Distribución
Palabras clave: Power system asset management, machine learning, artificial intelligence, unsupervised learning, reinforcement learning, health indexing, predictive maintenance, open-source toolbox, data quality, feature weighting, multi-dimensional assessment, QLearning, Self-Organizing Maps, clustering, optimization, transmission systems, distribution systems.
Cita:
G.L. Rajora, "An Intelligent Environment for Asset Management in Power Grids Based on Machine Learning Models", Tesis Doctoral, Universidad Pontificia Comillas, Madrid, España, 2026.
