Thesis
Application of machine learning for power system frequency stability : security constrained optimisation and adaptive emergency control
- Creator
- Rights statement
- Awarding institution
- University of Strathclyde
- Date of award
- 2025
- Thesis identifier
- T17495
- Person Identifier (Local)
- 202168023
- Qualification Level
- Qualification Name
- Department, School or Faculty
- Abstract
- This thesis develops a novel set of Machine Learning (ML)-based techniques for frequency stability management in modern power systems with high penetration of Converter-Interfaced Generation (CIG). Specifically, the techniques aim to enhance situational awareness by capturing detailed frequency dynamics, which conventional approaches struggle to do, and ensure frequency stability through preventive actions - by introducing ML-based constraints into optimisation models to account for the detailed dynamics - and corrective actions - by implementing agentic ML-based adaptive load shedding, focusing on locational aspects and optimising where, when and how much load to shed. Adapted for both online and offline applications, these techniques equip operators with advanced decision-making tools. Traditionally, system operators use detailed analytical expressions to manage and ensure secure operation. This approach captures and represents the frequency dynamics of the system by solving the Differential Algebraic Equations (DAEs) of the network. Furthermore, operators usually perform offline or time domain simulations (TDS) analyses on a selected set of scenarios, including expected worst-case scenarios. However, these approaches have significant computational requirements, making them typically suitable only for offline applications with a limited number of scenarios. Even when evaluating such a limited set of scenarios, identifying the most critical ones a priori has become increasingly difficult. As power systems become increasingly complex and uncertain, the challenge intensifies, with the growing number of candidate scenarios creating a significant computational hindrance. Consequently, operators often need to over-secure the system, which comes at a cost and hinders widespread adoption of clean energy to achieve net-zero targets. Moreover, numerical approaches - simplified alternative models for fast screening - are increasingly facing accuracy issues due to the increasing complexity of the system, potentially introducing significant errors, which can lead to unforeseen instability. In response to these challenges, this research proposes an ML-based approach to ensure frequency stability in modern low-inertia networks. ML models are capable of accurately learning complex system relationships using simulation or observational data used during training. As a result, they can predict the system response almost instantly, because solving the network’s DAEs is no longer required to make predictions. Such an approach is especially suitable for real-time or near-real-time applications, providing quick scenario screenings of detailed dynamics, where TDS would be computationally prohibitive. Ultimately, operators are provided with advanced decision-making tools, enabling them to more accurately and efficiently manage frequency stability while imposing minimal computational overhead.
- Advisor / supervisor
- Papadopoulos, Panagiotis
- Bukhsh, Waqquas
- Resource Type
- DOI
- Funder
Relations
Items
| Thumbnail | Title | Date Uploaded | Visibility | Actions |
|---|---|---|---|---|
|
|
PDF of thesis T17495 | 2025-10-24 | Public | Download |