Traffic Flow Prediction Model Based on Neural Ordinary Differential Memory Network
Traffic flow prediction is a typical task in multivariate spatio-temporal prediction and an important part of intelligent transport systems.However,existing models seldom pay attention to shared patterns among different roads in the traffic network.Most of mainstream models are implemented based on Graph Neural Network(GNN),which are subject to transition smoothing phenomenon,i.e.,the representations in the neighborhood graph tend to be similar,as the number of layers increases.In order to solve above problems,a Neural Ordinary Differential Memory Network(NODEMN)is proposed,which utilizes pattern memory units to retain salient features in spatio-temporal data for pattern matching.In addition,the over-smoothing problem that occurs in deep training is improved by using neu-ral differential equations.NODEMN conducts a large number of experiments on real datasets.Experimental results show that the NODEMN model has a significant advantage in prediction performance compared to the baseline methods,with an average reduction of 4.09%in the Mean Absolute Percentage Error(MAPE),Mean Absolute Error(MAE)by 3.38%,and Root Mean Square Error(RMSE)by 2.49%on three datasets.
multivariate time series predictionspatio-temporal databaseGNNmachine learningtraffic prediction
国家科技期刊平台
NSTL
万方数据
