Spatiotemporal Taxi Demand Prediction with Novel Clustering and Demand Disaggregation Framework

Accurate taxi demand forecasting is essential for optimizing urban transportation services and resource allocation. However, spatiotemporal variations, external factors, and heterogeneous demand patterns make accurate prediction challenging. Unlike existing works, this study introduces a novel non-neighbor clustering approach to improve taxi demand prediction by grouping taxi zones based on demand correlation rather than geographic proximity. The proposed clustering method integrates Kernel PCA and K-Means algorithm. Different predictors such as Random Forest, XGBoost, AdaBoost, and LSTM are compared with and without the proposed clustering method. To improve the training performance of the predictors, a scenario reduction method is applied. Additionally, a demand disaggregation framework is proposed to break down cluster-level prediction to the taxi zone-level. Among the evaluated models, Random Forest achieved the best results for the taxi zone-level prediction with an accuracy of approximately 80%. We provide the comparison with benchmark work to highlight the performance of the proposed model.