Abstract
This paper concerns predictions of freeway traffic flow under non-recurrent events using multivariate machine learning models, including the multilayer perceptron network and the one-dimensional CNN long short-term memory network. The machine learning architectures and loss functions for training neural networks are presented. The study region is a portion of the Kwinana Freeway northbound in Perth, Western Australia. The study dataset, obtained by matching the timestamp of all available data, has various features, including traffic volume (flow rate), speed, density and road incident. Using the root mean squared error and mean absolute error, results from the two learning models are compared to the baseline model to determine the suitable model for traffic prediction under non-recurrent events.