Physics–Data fusion–based traceable frequency trajectory prediction method for power systems

• Propose a physics-data fusion dual-network collaborative method for power system frequency prediction. • Embedding the SFR model as a physics-based loss term within the neural network reduces the model’s dependence on training data. • Using model-output physically meaningful parameters to reconstruct the frequency response enhances the model's physical traceability. • Combining PINN with LSTM enhances the model’s ability to capture temporal features. With the increasing integration of high-proportion renewable energy, the frequency response characteristics of power systems have become increasingly complex, making accurate frequency trajectory prediction critical for maintaining system stability and supporting operational decisions. To address the issues that existing neural-network-based frequency trajectory prediction methods exhibit a high dependence on training data and merely output the prediction results, which fails to establish traceable correspondence with system physical parameters and thus leads to insufficient physical credibility of the forecasts, this paper proposes a physics-informed and data-driven traceable prediction model. First, the third-order System Frequency Response (SFR) model is embedded into the loss function to constrain the neural network to follow physical laws, thereby reducing its dependence on training data. Then, a dual neural network architecture is employed to predict the SFR parameters and the imbalance power separately, and subsequently reconstructs the frequency curve to achieve post‑disturbance frequency trajectory prediction. This allows the prediction results to be mapped to key system physical parameters, thereby enhancing the physical traceability of the outcomes. Finally, the proposed method is validated on an improved 10-machine 39-bus New England system and CSEE-FS system. Results demonstrate that the model outperforms conventional approaches in terms of prediction accuracy and generalization capability.