Frequency Behaviour Prediction Based on Machine Learning for Future Low-Inertia Power Systems

Power systems worldwide are experiencing massive transformation with a large amount of renewable generation being integrated to replace fossil fuel-based synchronous generation. The rapid increase of renewables will significantly reduce the system inertia, leading to frequency changing faster during disturbances. Furthermore, compared to conventional Synchronous Generator (SG)-dominated systems, a more diverse types of resources (e.g., battery storage, demand side response, combined heat and power, etc.) are increasingly used for providing frequency support. As a result, frequency behavior in future power systems is expected to be more volatile and uncertain, resulting in significant challenges for frequency control. This paper aims at addressing the frequency control challenges in future low-inertia systems via providing an accurate prediction of frequency nadir and frequency trajectory during power imbalance events with the aid of machine learning techniques. The proposed method, based on Random Forest and Long Short-Term Memory Recurrent Neural Networks, can be used for predicting frequency behavior, thus supporting the effective design, schedule and dispatch of frequency responses. Case studies are presented, which demonstrate the proposed method is highly accurate, while being light weighted, which offers a promising tool to support future frequency solution design and dispatch.