STDACN: a Spatiotemporal Prediction Framework based on Dynamic and Adaptive Convolution Networks

Wei Yu, Jun Wang, Lei Wang, Nannan Wu, Guangquan Xu, Xiaoming Li

Keywords: Spatiotemporal Prediction; High-Aware Dependence; Dynamic TCN; Adaptive GCN.

TL;DR: A spatiotemporal prediction framework based on dynamic and adaptive convolution networks

Abstract: With the rapid advancement of sensor technologies, analyzing and modeling large spatiotemporal datasets has become crucial, enabling system state predictions for intelligent transportation, urban planning, public safety, and environmental protection. Current models—statistical, classical deep learning (e.g., TCN, GCN), and large-scale methods—struggle with noise, complexity, high dimensionality, and dynamics, with static TCN/GCN structures limiting performance and large models facing high computational costs, keeping classical methods relevant. This paper proposes a \underline{s}patio\underline{t}emporal prediction framework based on \underline{d}ynamic and \underline{a}daptive \underline{c}onvolution \underline{n}etworks (STDACN), which overcomes weight-sharing limits, featuring a high-order gated TCN with recursive causality to capture temporal dependencies and an adaptive GCN for spatial topologies, boosting efficiency and generalization. Excelling in traffic, weather, and population predictions across varied scales, STDACN offers a simple yet innovative path for classical deep learning in complex spatiotemporal modeling.

Primary Area: learning on time series and dynamical systems

Submission Number: 7048