SAMSGL: Series-aligned multi-scale graph learning for spatiotemporal forecasting.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Spatiotemporal forecasting is complex due to propagation dynamics and high-dimensional node interactions.
• Existing graph-based networks struggle with time delays and multi-scale interactions, limiting forecasting performance.

Purpose of the Study:

• To introduce the Series-Aligned Multi-Scale Graph Learning (SAMSGL) framework to enhance spatiotemporal forecasting.
• To address limitations in modeling time delays and multi-scale interactions in graph-based forecasting.

Main Methods:

• Developed a series-aligned graph convolution layer to aggregate non-delayed graph signals and mitigate time delay influence.
• Proposed a multi-scale graph learning architecture with global and local graph structures.
• Integrated graph-fully connected (Graph-FC) blocks to fuse spatial and temporal information.

Main Results:

• The SAMSGL framework demonstrated superior performance in meteorological and traffic forecasting experiments.
• The series-aligned convolution effectively handled time delays, improving prediction accuracy.
• Multi-scale graph learning captured both global and local spatiotemporal interactions.

Conclusions:

• SAMSGL framework significantly enhances spatiotemporal forecasting accuracy.
• The proposed methods effectively address time delays and multi-scale interactions in graph-based forecasting.
• SAMSGL shows promise for real-world applications in traffic and weather prediction.