Demand Forecasting for Emergency Supplies in Public Health Emergencies via Multimodal Semantic Alignment and Joint Trend–Fluctuation Modeling

Demand forecasting for emergency supplies is critical for public health emergency response, but demand sequences are strongly affected by external factors such as epidemic progression, traffic control, medical resource adjustments, and weather conditions, leading to stage-wise drift, short-term surges, and recovery-stage declines. Existing methods often rely on historical demand sequences or use external texts as sample-level auxiliary prompts, making it difficult to distinguish the semantic attributes, temporal granularity, and forecasting roles of different information sources. To address these limitations, we propose multimodal semantic alignment and trend–fluctuation forecasting (MATF). Following the information availability constraint in real forecasting processes, MATF organizes external information into four structured text fields, namely static context, observation-window events, historical constraints, and forecasting horizon prompts, and combines them with historical demand sequences to construct rolling forecasting samples. Methodologically, MATF uses Multimodal Input Encoding and Adaptation (MIEA) to preserve field boundaries and granularity differences, Context-Aware Event-to-Time Alignment (CETA) to align event semantics with historical time steps, and a Trend–Fluctuation Forecaster (TFF) to jointly model low-frequency trend evolution and semantic residual corrections. Experiments on real-world waybill-derived emergency logistics demand data and external text data from Wuhan show that MATF achieves lower forecasting errors than numerical-only baselines, text-enhanced baselines using the same information inputs, and ablation variants. Compared with AutoTimes, the best-performing baseline under identical information inputs, MATF reduces mean absolute error (MAE), root mean square error (RMSE), and symmetric mean absolute percentage error (sMAPE) by 9.9%, 9.2%, and 5.4%, respectively. Further ablation and sensitivity analyses support the effectiveness of structured text organization, event-to-time alignment, and joint trend–fluctuation modeling.