AI-Powered Crowd Dynamics Forecasting: A Large Language Model Framework for Public Safety and Decision Intelligence

Mass public gatherings, including festivals, sporting events, religious congregations, political rallies, and urban celebrations, create rapidly changing crowd patterns that challenge public safety authorities. Conventional forecasting approaches often depend on historical mobility or structured numerical data and have limited ability to incorporate event schedules, emergency alerts, and public sentiment. This paper proposes an AI-Powered Crowd Dynamics Forecasting framework that combines spatial–temporal learning with Large Language Model (LLM)-driven contextual reasoning. Historical mobility trajectories, GPS traces, event metadata, traffic information, meteorological conditions, social media streams, and Point-of-Interest data are integrated through a multimodal data-fusion pipeline. A graph-based mobility representation and Transformer temporal encoder learn spatial–temporal patterns, while the LLM interprets contextual information and supports semantic reasoning. The framework forecasts crowd density, pedestrian flow, congestion hotspots, and movement trajectories, followed by risk assessment and explainable recommendations for resource allocation, evacuation planning, traffic rerouting, and crowd diversion. The reported evaluation achieves MAE 4.18, RMSE 5.37, MAPE 6.84%, accuracy 97.8%, precision 97.4%, recall 97.1%, F1-score 97.2%, and AUC 0.989, demonstrating the intended value of multimodal contextual intelligence for proactive crowd management.