Trip-based prediction of hybrid electric vehicles velocity using artificial neural networks

In this paper, a high-performance Long Short-Term Memory (LSTM) neural network vehicle velocity predictor considering the case of countries with no vehicle to infrastructure or vehicle to vehicle data available. This fact restricts the amount of information that can be used for the network training process. The study takes into consideration the computational complexity of the developed predictor since it will ultimately be implemented as part of a real-time car controller. Two real-world driving cycles from developed and developing countries were collected from multiple drivers in order to make sure that the created datasets cover multiple driving patterns and scenarios. The considered trips include multiple driving conditions such as a highway, urban road, and intersections. Two architectures of time series prediction models are evaluated: the Non-linear AutoRegressive with eXogenous inputs (NARX) and LSTM neural networks. The proposed paper also explores the possibility of expanding the features of the networks beyond technical inputs to tackle macro-features such as the date, time of day, holiday etc., in order to test their effect on the overall prediction as well as the computational efficiency of the proposed velocity predictor. Results show that the LSTM model outperforms the NARX model and accurately predicts multi-step ahead vehicle velocity under various weather and traffic conditions while maintaining a low computational complexity.