Abstract
The integration of a high proportion of renewable energy has significantly reduced the grid inertia level and markedly increased the risk of transient frequency instability in power systems. Meanwhile, the large-scale integration of diverse heterogeneous resources-such as wind power, photovoltaics, energy storage, and high voltage direct current (HVDC) transmission systems-has considerably enriched the portfolio of frequency regulation assets in modern power grids. However, the marked disparities in the dynamic response characteristics and actuation speeds among these resources introduce significant nonlinearity and high-dimensional complexity into the system's transient frequency behavior. As a result, conventional methods face considerable challenges in achieving accurate and timely prediction of such responses. However, the substantial differences in the frequency regulation characteristics and response speeds of these resources have led to a highly nonlinear and high-dimensional complex transient frequency response process, which is difficult to accurately and rapidly predict using traditional methods. To address this challenge, this paper proposes an online prediction method for transient frequency response that deeply integrates physical principles with data-driven approaches. First, a frequency dynamic response analysis model incorporating the frequency regulation characteristics of multiple resource types is constructed based on the Single-Machine Equivalent (SME) method, which is used to extract key features of the post-fault transient frequency response. Subsequently, information entropy theory is introduced to quantify the informational contribution of each physical feature, enabling the adaptive weighted fusion of physical frequency response features and Wide-Area Measurement System (WAMS) data. Finally, a physics-guided machine learning framework is proposed, in which the weighted physical features and the complete frequency curve predicted by the physical model are jointly embedded into the prediction process. An MLP-GRU-Attention model is designed as the data-driven predictor for frequency response. A physical consistency constraint is incorporated into the loss function to ensure that predictions strictly adhere to physical laws, thereby enhancing the accuracy and reliability of the transient frequency prediction model. Case studies based on the modified IEEE 39-bus system demonstrate that the proposed method significantly outperforms traditional data-driven approaches in terms of prediction accuracy, generalization capability under small-sample conditions, and noise immunity. This provides a new avenue for online frequency security awareness in renewable-integrated power systems with multiple heterogeneous frequency regulation resources.