Multi-dimensional origin-destination freight flow prediction via a hybrid multi-graph convolutional neural networks based model

Abstract

With the rapid development of eCommerce, freight plays an increasingly important role in contemporary society. The global pandemic further highlights the significance of the freight system. Nowadays, people gradually place more and more reliance and expectations on online shopping, making freight activities more complex. As an important and complicated topic, most existing freight flow prediction studies are limited. For example, the studies either focus on linear temporal dependencies or focus on spatial-temporal dependencies. A multi-dimensional freight flow prediction model is needed to disentangle the freight flow variation pattern better. Additionally, many new freight alternatives emerge thanks to the development of information technology and artificial intelligence. Massive freight data is available, enabling the application of advanced machine learning and deep learning (DL) models. In order to provide a comprehensive and accurate model specialized in freight flow, a multi-graph convolutional based DL model framework is proposed to predict freight flow between origins and destinations considering impacts from the dimensions of temporal, spatial, and socioeconomic features. The DL model consists of three main parts: a long-short term memory module to capture temporal dependencies, a two-dimensional graph convolutional module to capture spatial and socioeconomic features, and a graph fusion module to integrate impacts from multiple dimensions. Following this, an input-output (I-O) table is disaggregated into a lower spatial level according to the distribution of predicted freight flow among different commodities. An empirical application of the proposed model framework is analyzed using a real-world dataset collected from a leading crowdsourcing freight company in China. The case study predicts the crowdsourcing freight flow of each commodity type between each pair of cities in China, including dependencies of time, geographic location, and city characteristics. China’s I-O table is further disaggregated into a city-level I-O table to reveal the inter-commodity dependencies between some city pairs. Unique features of the studied dataset are discussed, and the prediction power of the proposed model is demonstrated. As a result, this dissertation contributes to freight flow studies by offering an end-to-end multi-dimensional DL model framework to predict freight flow OD matrix, and an efficient alternative to disaggregate I-O table to reveal inter-commodity dependencies between smaller spatial units.