Flight delay propagation in aviation network poses a significant challenge to air traffic management and airline operations. Existing approaches often fall short in accurately predicting delay propagation due to their failure to comprehensively model the intricate dynamic interactions between spatial and temporal factors in the aviation network. To address this gap, this study proposes a Spatial-Temporal Dynamic Interactive Graph Neural Network (STDIGNN) framework. First, it integrates the complex network properties (e.g., clustering coefficients and centrality) into a diffusion graph convolution module to exploit spatial dependencies, and employs multiple multilayer perceptrons to capture temporal correlations. Then, it builds a spatial–temporal bi-directional interactive learning mechanism to reinforce the dynamic correlations between time and space, facilitating a more profound understanding and more precise prediction of the delay propagation. Extensive experiments on two large-scale real-world datasets from China and the U.S. show that STDIGNN outperformed state-of-the-art baselines, reducing the prediction mean absolute error in average by 43.85 % (arrival) and 46.92 % (departure) in China, 14.7 % (arrival) and 19.00 % (departure) in the U.S., and achieved high precision of 5.374 min in China and 4.603 min in the U.S. for 6-hour predictions. This study provides valuable insights and practical solutions for optimizing air transportation systems. The implementation code is available at https://github.com/sculiang/FlightDelayPrediction_STDIGNN.