Spatiotemporal heterogeneity-aware ridership prediThe calculation results are shction with MGWR-informed attention on CNN-BiLSTM

Accurate taxi ridership forecasting plays a vital role in enhancing passenger experience and mitigating the mismatch between taxi supply and demand. However, most existing methods approach the problem from a city-wide perspective, overlooking the spatiotemporal heterogeneity inherent in the built environment. Additionally, traditional forecasting models tend to emphasize temporal dynamics while neglecting spatially varying environmental features, resulting in limited adaptability and predictive accuracy. To address these challenges, this study proposes a novel framework that integrates Multiscale Geographically Weighted Regression (MGWR) as an attention mechanism into a deep learning model to capture the complex interplay between urban form and ridership. MGWR is first used to uncover localized, multiscale dependencies between taxi demand and built environment variables, effectively characterizing regional heterogeneity. These insights are then embedded as spatial attention signals into a CNN-BiLSTM model, enhancing its ability to fuse spatial and temporal patterns for improved ridership forecasting. Extensive experiments on real-world taxi data from Chengdu, China, demonstrate the effectiveness of the proposed method. The results reveal pronounced spatiotemporal heterogeneity: factors such as shopping centers, financial facilities, and bus stops exert stronger effects in central urban areas but show diminishing influence in suburban areas. Temporally, shopping centers contribute more to weekend ridership, while enterprises and residential areas play a greater role during workday commuting peaks. By incorporating these variations, the CNN-BiLSTM-Attention model achieves superior predictive accuracy, with an adjusted R² of 0.91 and much lower than the baseline error. These findings offer practical insights for urban transportation planning and support data-driven decisions for optimizing taxi dispatch and infrastructure design.