A universal geography neural network for mobility flow prediction in planning scenarios

Abstract

This study primarily focuses on generating mobility flow in regions and cities, which plays an important role in urban planning and management. The majority of existing mobility flow models, including conventional statistical models and deep learning-based models, are heavily dependent on historical data to predict future mobility flows. The application of these models poses significant challenges in the planning and construction of emerging cities and regions, particularly in developing countries experiencing swift urbanization. These challenges are exacerbated by a dearth of historical data and rapid shifts in mobility patterns. Consequently, the scenario necessitates a mobility flow generation model capable of generating flows without historical data. This study introduces the universal geography neural network, an algorithm designed to glean potential patterns in human mobility across diverse cities and temporal spans. This is achieved through the analysis of substantial quantities of location-based data, resulting in the generation of mobility flows within a city. Our experiment, designed to extract various features and generate fine-grained mobility flows in the testing set, outperforms both traditional models and state-of-the-art deep learning models. Moreover, our model has proven capable of generating reliable results across various time periods and grid areas.