Machine Learning Pipelines for Intelligent Alarm Management in Power Systems

Abstract:
The effective management of alarms in power systems faces significant challenges as operators are increasingly overwhelmed by the large volume of events and alarms, which many times require no action. This issue can be confirmed by some recent cases from European system operators, where hundreds of thousands of SCADA events are recorded daily, potentially compromising operators' ability to identify and respond to critical alarms. To address this challenge, this paper proposes the application of machine learning techniques for intelligent alarm management, developed in collaboration with a transmission system operator. We identify five key machine learning use cases for intelligent alarm management: alarm text clustering, alarm flood clustering, similar alarm identification, outlier detection, and alarm type suggestion. For each use case, we introduce a concrete pipeline, which is designed with an operator-centric approach, enabling con-tinuous system improvement through active operator interaction and feedback integration. The paper provides compre-hensive implementation guidelines for each pipeline, including data requirements and retraining considerations, and val-idates three approaches using real operational data from a transmission system operator, demonstrating its practical fea-sibility in real-world scenarios.