Abstract: Existing ride-hailing demand forecasting methods suffer significant accuracy degradation during non-routine traffic events due to insufficient global multimodal pattern capturing and inadequate abnormal flow modeling. To address this issue, a data augmentation method was proposed, integrating domain-adaptive fine-tuning of Llama3 with semantic event fusion. A quantitative representation system was constructed, incorporating event importance and congestion impact coefficients, converting unstructured event descriptions into structured numerical features. A dual-path InformerGRU （Informer with Gated Recurrent Unit）architecture was designed, combining the Informer encoder’s long-sequence global modeling capability with the local sequence dynamic capturing advantages of bidirectional Gated Recurrent Units (GRU). Multi-scale features are intelligently aggregated through an adaptive fusion layer. Real-world experimental results demonstrate that the proposed data augmentation method reduces mean squared error (MSE) by 71.44% in abnormal traffic flow scenarios, with enhanced peak prediction. Compared to Informer, InformerGRU model achieves 4.33% lower mean absolute error (MAE) and improves accuracy by 2.68%. In typical traffic control scenarios, the proposed model reduces MAE by 49.33%.