Fluid dynamics simulation remains a critical challenge in computer graphics and engineering, where traditional methods like Smoothed Particle Hydrodynamics face limitations in computational efficiency, global feature modeling, and handling complex boundary conditions. This paper proposes a hybrid architecture that integrates continuous convolution networks with multi-head weighted self-attention mechanisms to address these challenges. The continuous convolution layer mimics SPH’s spatial weighting properties, enabling efficient local interaction modeling, while the attention mechanism captures long-range particle correlations and mitigates the quadratic complexity of traditional attention through grid-based pooling and spatial encoding. Additionally, relative positional encoding and adaptive pooling strategies enhance the physical plausibility and stability of simulations. The framework has attained good results in fluid dynamic prediction works and effectively balances accuracy and computational efficiency, offering a robust solution for real-time fluid simulations in complex environments.