Predicting residents'negative emotional responses at individual level during public emergencies based on LLM

Increasingly frequent public emergencies significantly affect residents' emotions. Quickly predicting shifts in residents' negative emotions during emergencies is key to improving response strategies. This study proposes a Qianfan-LLM framework to predict individual-level shifts in negative emotional intensity during emergencies with understandable explanations, using the 2015 Guangming landslide in Shenzhen as a case. The framework integrates multiple inputs: descriptions of public emergencies, residents' pre-event negative emotional intensity, built environment, and socioeconomic features. The fine-tuned Qianfan-LLM achieved 74.63 % accuracy, outperforming traditional models (e.g., logistic regression, random forest) by up to 22.02 %, and two text embedding models Qwen3-Embedding-4B and Linq-Embed-Mistral, by 4.23 % and 4.95 % respectively. By leveraging LLMs, response strategies can shift from reactive to proactive, enabling personalized interventions for emotionally vulnerable residents and facilitating more targeted emergency response plans. This study explores applying LLM for individual-level emotional prediction during public emergencies, providing more accurate and interpretable decision support for emotional intervention strategies.