Spatio-Temporal Graph Network-Based Traffic Forecasting Method

Traffic forecasting is widely recognized as a fundamental component of smart city development; however, classic models struggle to manage long-term time series data effectively. In other words, existing models rely on multiple algorithm modules, which significantly increase computational costs and, in turn, reduce model accuracy. Furthermore, any new input dataset requires retraining the model from scratch. To address these challenges, this study proposes a novel approach, where a long-sequence embedding module and an encoder-decoder structure integrating a temporal attention layer and a graph convolution operator are introduced during the pretraining section. In the fine-tuning section, the learned parameters are frozen, and the decoder is replaced with a prediction header that incorporates a meta-learning module. This approach enables the integration of both long-term and short-term time series data while facilitating training tailored to specific target tasks. The method has demonstrated strong performance on real-world datasets, providing clear evidence of its effectiveness in addressing traffic forecasting challenges.