Content area
Abstract
Traffic flow prediction, as a key link in the intelligent transportation system, assumes the important role of efficiently guiding the traffic flow, evacuating, congestion, reducing traffic accidents, and so on. However, due to the complex spatial and temporal correlation of traffic flow data, it faces the problem of inaccurate short-term prediction. In this paper, we adopt retentive network (RETNET) as the infrastructure of large-scale language model, which is similar to the Transformer model, but combines the recursive advantage of RNN to realize the efficient operation of parallelism and recursion. The RETNET model also handles the long sequences of information by stacking the same modules, but the difference is that it introduces multi-scale retention module (MSR) instead of the multi-head attention mechanism in the Transformer model, and adopts the chunked recursive approach to reduce the inference cost and improve the decoding throughput. Transformer model, and adopts chunked recursive parallel processing to reduce the inference cost and improve the decoding throughput. It is then combined with a Chebyshev graph convolutional neural network to utilize the spatial correlation of graph nodes to aggregate and update the features of road intersection nodes. The temporal and spatial information of traffic flow data is fully utilized by the combined spatial and temporal feature extraction, which improves the accuracy and robustness of traffic flow prediction.