Real-time and accurate prediction of the long-term behavior of dynamic systems is crucial for identifying risks during unexpected events, while computational efficiency is significantly influenced by the scale of the dynamic system. However, existing neural network models mainly focus on optimizing network structures to improve accuracy, neglecting computational efficiency. To address this issue, we propose regional graph representation, which reduces the scale of the graph structure by merging nodes into region, extracting topological information through graph convolution or lightweight convolution modules, and restoring the regions via fine-grained reconstruction. Notably, this method is adaptable to all graph-based models. Meanwhile, we introduce a sparse time-aware expert module, which selects experts for processing different scale information through a dynamic sparse selection mechanism, enabling multi-scale modeling of temporal information. The architecture we achieve an optimal balance between speed and prediction accuracy, providing a practical solution for real-time forecasting.
Peng et al. (Thu,) studied this question.