Velocity Prediction Based on LSTM: Impact of Different Input Settings on Prediction Performance

Abstract

A large number of intelligent driving systems that rely on the velocity prediction of the host vehicle or other road users are constantly emerging. With the development of nonparametric methods, artificial neural network has been widely employed in the predictive task in the past few years and a significant representative is long short-term neural network (LSTM NN). One of the noteworthy advantages of LSTM is its outstanding ability to overcome the issue of back-propagated error decay and thus demonstrated excellent effect of time series forecasting with long-term dependence. At present LSTM has been deeply researched and has been proved to be a very effective manner to capture nonlinear velocity dynamics. In this paper, various input settings are introduced to study how different variables affect the predictive performance of LSTM. Historical acceleration, the velocity of preceding vehicle, and the distance between adjacent cars are used as supplementary information that input to the model for velocity prediction and the application of real data validated that the predictive performance of the model varies with the input variables. The results show that the inclusion of the velocity of preceding vehicle help to enhance the performance of the model best overall.