Abstract: With the increase of uncertainty factors such as variability of renewable energy and load randomness in power systems, especially in N−1 contingency scenarios, efficient large-scale repetitive power flow calculation is becoming increasingly crucial for real-time security analysis. However, traditional power flow calculation methods based on physical mechanisms have higher computational costs and slower speeds, which can not meet the real-time risk assessment requirements. Data-driven power flow calculation methods have faster speed but rely heavily on data quality, and the prediction results need to be more consistent with physical mechanisms, making it challenging to apply to actual industrial scenarios. To address these issues, this paper introduces power system domain knowledge into data-driven models by constructing a deep learning model that complies with physical constraints, thereby improving the model's performance. It embeds the power system topology structure and physical formulas into the deep neural network structure through a gated mechanism and regularization strategy, enabling the model to adapt to changes in network topology in N−1 contingency scenarios. This paper conducts simulation experiments using the IEEE 14-node and IEEE 39-node systems with new energy access, investigating the model's performance in conventional and N−1 fault scenarios. The experimental results show that the proposed method has improved accuracy and compliance with physical constraints compared to traditional deep learning power flow calculation methods, and can effectively evaluate the system's operating state under different fault conditions, verifying the effectiveness of the proposed method.