Thesis
Enhanced frequency control in low-inertia power systems using learning-based estimation and prediction
- Creator
- Rights statement
- Awarding institution
- University of Strathclyde
- Date of award
- 2026
- Thesis identifier
- T18049
- Person Identifier (Local)
- 202273421
- Qualification Level
- Qualification Name
- Department, School or Faculty
- Abstract
- The growing penetration of Inverter-Based Resources (IBRs) has significantly transformed power system dynamics, leading to reduced system inertia and increasingly complex system behaviour. Consequently, traditional frequency control practices based on offline studies and relatively fixed contingency assumptions are becoming increasingly inadequate for maintaining frequency stability under rapidly changing system conditions. In particular, existing approaches that rely on analytical models and offline simulations are limited in their ability to capture complex and volatile system dynamics and to adapt in a timely manner. Furthermore, frequency control actions are predominantly reactive, typically being triggered only after frequency deviations have already occurred. While this is generally adequate for moderate frequency events, it may prove insufficient during major power imbalance disturbances. Their effectiveness further depends on fast and accurate measurements of key system parameters, particularly inertia, Rate of Change of Frequency (RoCoF), and power imbalance size, which are difficult to estimate reliably in low-inertia power systems. Recent advances in learning-based techniques have created new opportunities for addressing a range of frequency stability challenges arising from the large-scale integration of IBRs. However, existing research has largely focused on direct control implementations or narrowly scoped prediction tasks, with comparatively limited attention given to the development of proactive, situation-aware, and practically deployable decision-support methodologies. As a result, a clear gap remains in translating learning-based modelling advances into effective real-time operational support for low-inertia power systems. This thesis addresses this gap by leveraging the real-time estimation and prediction capabilities of learning-based methods to enhance frequency control in low-inertia power systems, explicitly complementing and strengthening existing control and operational practices rather than replacing them. Using realistic data and tools available to system operators, the proposed approaches enhance frequency control across multiple stages by improving risk-informed scheduling, disturbance estimation, and adaptive emergency response. In this thesis, frequency control is conceptualised as three interconnected stages, i.e. pre-event scheduling, the power imbalance event stage, and post-event frequency containment. Learning-based techniques are applied across all three stages to improve the accuracy and effectiveness of frequency control, thereby establishing a more comprehensive and practical estimation and prediction capabilities for future power systems with low-inertia and increasingly complex operating conditions. In the pre event scheduling stage, this work proposes a risk-aware prediction component to capture probabilistic frequency nadir behaviour across diverse operating conditions. Instead of relying on deterministic or single-point nadir predictions, the proposed approach explicitly quantifies prediction uncertainties, particularly those arising from real-time inertia measurements. These uncertainty-informed intervals are directly integrated into the optimisation process, which enables the model to ensure that scheduling decisions account for the most severe potential prediction errors. Consequently, the approach facilitates a balance between minimising procurement costs and maintaining stringent frequency stability requirements, while preserving security margins even when prediction accuracy is degraded. During the power imbalance event stage, which corresponds to the initial transient following a disturbance, the rapid and reliable estimation of RoCoF and power imbalance size is critical for supporting timely frequency control actions. A Discrete Wavelet Transform (DWT)-Support Vector Regression (SVR) RoCoF estimator is developed to balance estimation accuracy and computational speed under noisy and fast-varying measurement conditions. The estimator addresses limitations of conventional RoCoF measurement techniques, which are often sensitive to transients and measurement noise. In addition, an Extreme Gradient Boosting (XGBoost) power imbalance size estimation method is introduced to overcome the limitations of traditional swing equation-based approaches that rely on accurate system inertia information and reliable RoCoF measurements. The proposed method infers the active power imbalance directly from frequency measurements and system operating conditions. These estimation modules provide essential situational information during contingency events to support more effective subsequent frequency control decisions. To further safeguard the system during disturbances, the post-event frequency containment stage is addressed through the real-time prediction of frequency behaviour and an adaptive emergency control scheme it enables. Based on the predicted frequency behaviour and system state information, the controller adaptively adjusts load shedding thresholds and stages in real-time to contain frequency deviations, which avoid frequency decline beyond critical thresholds during severe disturbances. The three stages are linked through a progressive, learning-based decision-support structure across different time scales, i.e. pre-event scheduling is guided by probabilistic assessments of frequency nadir risk, disturbance estimation during the initial transient provides timely and detailed system state information, and post-event frequency containment relies on adaptive control decisions informed by predicted frequency behaviour. The proposed methodologies across all three stages are validated through extensive real-time simulations, Hardware-in-the-Loop (HiL) experiments and real world recorded frequency event data. The results demonstrate that learning-based estimation and prediction enhance frequency control performance in low-inertia power systems by improving scheduling robustness, real-time disturbance awareness, frequency nadir containment compared with conventional approaches.
- Advisor / supervisor
- Hong, Qiteng
- Resource Type
- DOI
Relations
Items
| Thumbnail | Title | Date Uploaded | Visibility | Actions |
|---|---|---|---|---|
|
|
PDF of thesis T18049 | 2026-06-10 | Public | Download |