Adaptive Denoising Spatio-Temporal Attention Network Fused With External Factors for Passenger Flow Prediction

Accurate passenger flow prediction is important in intelligent transportation systems (ITS). As is well known, introducing external factors, such as weather and air quality data, can enhance the representation of passenger flow characteristics, which has a certain positive impact on improving predictive performance. However, if there is no effective denoising, the more external factors introduced, the lower the predictive performance. Therefore, there are few studies that incorporate external factors into modelling. To this end, we propose an Adaptive Denoising Spatio-Temporal Attention Network (ADSTA-Net) that integrates external factors for passenger flow prediction. The core of the model is to fully consider the impact of external factors on passenger flow. In the initial stage of ADSTA-Net, multiple external factors are combined with passenger flow. Then, the adaptive learning parameter matrix (ALPM) and fast Fourier transform (FFT) are applied to perform adaptive denoising for the fused features at different times and locations. Finally, a simplified Graph Multi-Attention Network (GMAN) with only-one layer ST-Attention block is used to learn global spatio-temporal dependencies. Extensive experiments are conducted on two real-world passenger flow datasets, and the results demonstrate that ADSTA-Net has superior performance, particularly in making more accurate predictions under bad weather conditions.