MAD-GCN: Multi-attention Diffusion Graph Convolutional Network: A Novel Approach for Low-Altitude Traffic Flow Forecasting

Low-altitude traffic flow forecasting is the core task of urban management and Intelligent Transportation System (ITS). Due to the complexity and uncertainty of traffic patterns, capturing the spatial-temporal correlation of traffic patterns is a big challenge. Recent research methods mainly rely on the predefined adjacency matrix to simulate the traffic relationship, which leads to poor prediction effects. Meanwhile, if only the adaptive adjacency matrix is relied on, some inevitable relations between nodes may be ignored. To solve the above problems, this paper proposes a novel deep learning model: Multi-Attention Diffusion Graph Convo-lutional Network (MAD-GCN). The model consists of three main modules: 1) Spatial-Temporal Positional Embedding module, in which historical and predicted traffic flows are embedded in the form of spatial-temporal embedding vectors; 2) Spatial correlation module, in which dynamic node relationships are captured and spatial models are established; 3) Time correlation module, in which introduces multi-attention to capture a variety of temporal features in traffic data series. Our proposed model was compared with multiple baseline models on real datasets, and the results show that our MAD-GCN is optimal and can capture the spatial-temporal dependence well.