Taxi demand forecasting considering spatiotemporal heterogeneity in built environment by deep learning

Accurate forecasting of taxi demand enhances the customer travel experience and provides important support to alleviate the imbalance between taxi supply and demand. Existing research primarily conducts taxi demand prediction from an overall city perspective, overlooking spatiotemporal heterogeneity in the built environment. At the same time, traditional models struggle to capture spatiotemporal features, resulting in low accuracy and poor portability. This study investigates a taxi demand forecasting model that can consider the spatiotemporal heterogeneity of the built environment. This study first employs a Multiscale Geographically Weighted Regression (MGWR) model to analyse the spatiotemporal heterogeneity between taxi demand and the built environment, then develops a CNN-BiLSTM-Attention model to integrate these spatiotemporal features for accurate demand forecasting, and finally validates the model's effectiveness through extensive real-world experiments in Chengdu, China. The model in this paper has high prediction accuracy, with MAE, MAPE, and Adjusted R-Square indicators of 10.45, 16.36%, and 0.90, respectively. The results contribute to a better understanding of the relationship between the built environment and taxi demand, supporting urban managers in adjusting the supply and demand allocation of taxis within the city.