Junction Traffic Prediction, Using Adjacent Junction Traffic Data, Based on Neural Networks

The paper discusses the problem of substitution of traffic data, used for prediction of traffic flows at a junction, with data from an adjacent junction. Such a case arises when the measuring resources at the junction malfunction. Neural networks based approach is used for forecasting traffic flows. Solutions incorporating a multilayer perceptron (MLP) network, a cascade forward network (CFN) and a deep learning network (DLN) with autoencoders are used for evaluating the prediction performance. The elaborated designs are validated using a data set of traffic flow measurements comprising over 15 thousand measurements collected in a period of over six months. Results prove that substituting data from an adjacent junction is justified for predicting traffic flows in case of malfunctioning measuring resources.