An Surrogate Model for Predicting Lightning Transient Waveforms by Integrating Physical Constraints and Attention Mechanism

Accurate modelling of lightning transient processes is crucial for power system protection design. However, the traditional electromagnetic transient numerical simulation is time-consuming and requires manual setting of simulation conditions, which is difficult to meet the demand of large-scale random data generation. Aiming at the randomness of lightning current waveforms, this paper proposes a deep learning surrogate model, Physics-informed Attention U-Net (PAU-Net), which integrates physical constraints and Attention Mechanism. The model is based on a one-dimensional U-Net, incorporates priori physical information such as tower location and lightning amplitude into the input via a conditional feature encoder, and designs a multi-component composite loss function containing peak, gradient, and smoothness to ensure the physical consistency of the prediction results. The experiments show that the prediction accuracy of PAU-Net is significantly improved compared with the baseline U-Net, with the Mean Absolute Error reduced by 18.5%, the Peak Relative Error controlled within 2.8%, and the computational efficiency improved by two orders of magnitude compared with the traditional simulation. The model not only predicts lightning transient waveforms with high accuracy, but also has the ability to generate large-scale physically consistent data, providing an efficient and scalable new method for lightning risk assessment and electromagnetic transient analysis.