Research And Implementation Of Traffic Forecasting With Spatio-Temporal Data Mining

Existing urban traffic management methods are insufficient to provide safe and convenient travel for pedestrians under complex road conditions.Therefore,intelligent transportation systems are needed to assist in the construction of smart cities.Traffic forecasting is one of the core research tasks of intelligent transportation systems,which aims to predict urban traffic conditions such as flow and speed,and future traffic conditions can help traffic management departments make decisions in advance to reduce traffic congestion and avoid traffic accidents.However,traffic flow or speed data is not only affected by spatial roads but also changes with time.Briefly,traffic data has spatio-temporal attributes,and traditional methods are hard to mine spatio-temporal information.Although the development of deep spatio-temporal data mining technology has brought progress to traffic forecasting,the challenges caused by the inconsistent spatiotemporal patterns of speed and flow are still ignored.According to the specific characteristics of traffic speed and flow data,this paper proposes two deep spatio-temporal data mining models for speed and flow forecasting.The main contents and innovations are as follows.For the traffic speed forecasting task,a deep spatio-temporal joint Transformer model,named Lastjomer,is proposed to solve the problem of incomplete spatio-temporal correlations mining.Lastjomer adopts an encoder-decoder architecture,and the encoder and the decoder are composed of a spatio-temporal joint attention mechanism to extract the cross-time spatial correlation and the cross-node temporal correlation.The spatio-temporal joint attention mechanism integrates the convolution and linear mechanism to enhance spatio-temporal local representation power and reduce the spatio-temporal complexity.Finally,the experimental results on two traffic speed datasets show that the performance and the model efficiency of Lastjomer are better than state-of-the-art traffic speed forecasting baselines,and it has good interpretability.For the traffic flow forecasting task,a deep spatio-temporal wavelet model,named STWave,is proposed to solve the problem of distribution shift of high-frequency components on the temporal dimension and unbalanced structural information on the spatial dimension.STWave uses a disentangled-fusion framework to model different frequency components separately to alleviate the influence of the distribution shift of highfrequency components.The disentangled part is completed by discrete wavelet transform,and the fusion part is an attention mechanism.The attention mechanism can adaptively choose correct forecasting results of high-frequency components.The spatio-temporal data mining part utilizes different temporal neural networks according to attributes of different frequency components and proposes a graph wavelet-based spectral graph attention network to extract local-global balanced spatial information.Besides,spectral graph attention network uses a query sampling strategy to reduce the complexity of the spatial dimension,which results in it can be applied to real-world scenarios.Finally,experimental results on four traffic flow datasets show that STWave achieves the best forecasting performance with low time-memory usage and good interpretability.