<p>Spatio-temporal dynamic modeling is a key research area with applications in weather forecasting, traffic flow prediction, and simulation of physical phenomena. However, traditional models face significant challenges in long-term prediction and extreme events. These challenges include maintaining accuracy and structural similarity over long time steps and ensuring robustness and precision during extreme events like fires and floods. This paper introduces a new spatio-temporal dynamic modeling method called the Spatio-Temporal Diffusion Network (STDN). The model combines traditional Convolutional Neural Networks with diffusion processes to improve performance in long-term prediction and extreme events. The model structure includes multi-scale convolution modules and diffusion modules. The multi-scale convolution module consists of an encoder, a temporal evolution module, and a decoder to extract and transform spatio-temporal features. The diffusion module optimizes feature representation progressively to enhance prediction accuracy. Experimental results show that STDN outperforms existing baseline models on multiple datasets, especially in long-term prediction and extreme events. On the Moving-MNIST dataset, STDN reduces the Mean Squared Error (MSE) by 46.5% compared to the second-best model. On the TaxiBJ dataset, it reduces MSE by 52.5%. Additionally, on the Navier–Stokes dataset, STDN maintains a high Structural Similarity Index and Peak Signal-to-Noise Ratio over long time steps, demonstrating its effective modeling of complex physical dynamics. In summary, by combining CNN and diffusion processes, STDN shows significant performance advantages in long-term prediction accuracy and handling extreme events. Experimental results validate its effectiveness in various spatio-temporal prediction tasks.</p>