Noise-immune Learning For Short-term Traffic Flow Forecasting

With the wide application of traffic sensors and the development of emerging sensor technology,traffic flow data has increased significantly.It is feasible and necessary to collect and analyze traffic flow data and predict short-term traffic flow through information methods.However,traffic flow is a real-time,completely nonlinear,high-dimensional,and non-stationary stochastic process.The elusive traffic flow pattern naturally contains noise caused by internal and external variations,including traffic accidents and extreme weather,so short-term traffic flow forecasting is still challenging.In the past decades,scholars have proposed straightforward and effective methods for predicting traffic flows under different traffic conditions,each with its respective merits and demerits.However,most studies on traffic flow forecasting are based on normal traffic conditions,and the interference of noise in traffic data to the model is rarely considered.To address the above issue,this study proposes two new short-term traffic flow prediction models based on noise immune under different original traffic data conditions.The main research contents are as follows:1)In order to utilize the temporal and spatial correlation contained in traffic flow data and suppress the noise contained in traffic flow data,a short-term traffic flow prediction model based on noise immune and temporal and spatial characteristics is proposed.Firstly,the wavelet thresholding denoising method is used to extract valid signals.Then the graph convolutional network(GCN)is used to capture the spatial dimension features of traffic flow data,followed by the long short-term memory network to extract valid signals.LSTM)captures dynamic changes in road traffic data to simulate time dependence.Experimental results reveal that the method yields more competitive results compared to previous models.2)Due to the noise caused by internal and external changes in the traffic flow signal and some essential features and heterogeneous correlations hidden in the raw traffic flow data,the further improvement of the model’s performance is restrained.To solve the above problems,we propose a novel noise-immune and attention-based multi-modal model for short-term traffic flow forecasting,which predicts traffic flow in a two-dimensional driven.In this case,it is reasonable to use a powerful and effective trunk to fully extract rich information to explore the implicit change mode of traffic flow and improve the model’s prediction performance.In addition,to improve the antiinterference ability to noise,a dynamic noise immune loss function based on the maximum entropy criterion is constructed.Aiming at verifying the validity of the method,extensive experiments are carried out on four real-world benchmark datasets to reveal the model’s superiority.Furthermore,additional ablation experiments are carried out to study the importance of each component.The main results show that this method can predict the future traffic flow more accurately.