Across temporal–spectral domains: Graph convolutional network with Fourier neural operator for spatiotemporal traffic flow prediction

Accurate prediction of traffic flow is essential for the development of intelligent transportation systems. Current approaches primarily concentrate on capturing the spatiotemporal dependencies within the data, often overlooking critical global characteristics of traffic flow present in the spectral domain. To address this limitation, the present study introduces a cross-domain fusion network (CDF-GFNet), which integrates a graph convolutional neural network (GCN) with a Fourier neural operator (FNO). The proposed framework integrates distinct spatiotemporal and FNO blocks to extract features from the spatiotemporal and spectral domains, respectively. It employs a cross-domain attention fusion mechanism to facilitate feature interaction, thereby establishing an interactive perception cross-domain representation learning framework for traffic flow prediction. This architectural design allows CDF-GFNet to more effectively leverage information derived from multiple data perspectives. Experimental evaluations conducted on the Ningde Expressway dataset from Fujian Province, China, demonstrate that CDF-GFNet exhibits superior capabilities in traffic flow feature extraction and real-time prediction, outperforming existing baseline models in predictive accuracy. The ablation experiment further demonstrated the effectiveness of spectral domain feature extraction. In addition, CDF-GFNet also shows favorable computational efficiency and parameter efficiency.