<p>Holiday periods often cause sharp fluctuations in online ride-hailing demand, creating significant challenges for capacity allocation and urban traffic management. This study proposes an attention-enhanced spatiotemporal Transformer model to predict peak demand during holidays, integrating temporal features, spatial patterns, and holiday-specific factors through an embedding mechanism. The model was evaluated using real-world ride-hailing datasets from major Chinese cities under various holiday scenarios. Experimental results show that, compared to baseline models such as ARIMA, SVM, and LSTM, the proposed model reduced RMSE by 15.2%, MAE by 12.4%, and MAPE by 14.8%, with an R² improvement of 9.6%. These improvements can potentially reduce average passenger waiting times by approximately 18% and lower idle driving rates by 12%, enhancing both service efficiency and sustainability. The framework is simulation-based, enabling flexible adaptation to diverse urban contexts, and is designed for integration into intelligent transportation systems for real-time traffic management. While the model demonstrates strong predictive performance, limitations include dependency on high-quality spatiotemporal data and reduced accuracy under extreme data sparsity. Future research will focus on improving robustness, extending the framework to multimodal transport integration, and exploring transfer learning for cross-city adaptation.</p>