Convolutional Spatio-Temporal Prediction Network with Trainable Positional Encoding

Existing spatio-temporal prediction networks that rely on recurrent neural networks face significant parallelization challenges, leading to high computational costs and prolonged training durations. In contrast, recurrent-free methods, particularly convolutional neural networks (CNNs), have attracted significant attention because of their lightweight models and straightforward architectures. Traditional CNNs, however, are limited by their inherent weight-sharing property, which imposes a strong inductive bias and restricts their ability to capture spatial heterogeneity. To overcome these challenges, we propose a novel trainable positional encoding convolutional neural network (TPE-CNN). TPE-CNN incorporates a positional encoding module to dynamically learn location-specific spatio-temporal correlations and temporal evolution patterns, addressing the inability of CNNs to handle absolute positional dependencies effectively. Additionally, we integrate large-kernel convolutions and inception modules, significantly enhancing global information awareness. This design enables the incorporation of global attention mechanisms while maintaining a low parameter count, achieving an optimal balance between performance and computational efficiency. To validate the effectiveness of TPE-CNN, we performed extensive experiments on three real-world datasets. Experimental results demonstrate that TPE-CNN delivers superior prediction accuracy while maintaining low computational complexity.