The analysis of online public opinion on public emergencies is of great significance to government decision-making and social governance. In this paper, a model for analyzing online public opinion on public emergencies is constructed, and a text clustering method combining sentence-level similarity calculation and sentiment analysis is proposed. Topic words are extracted by TF-IDF algorithm, clustering analysis is carried out by K-means and DBSCAN algorithm, and a plain Bayesian model fused with multiple sentiment lexicons is used to optimize the sentiment polarity classification. Taking the “Zhengzhou 720 Heavy Rainstorm Disaster” in 2021 as a case study, based on 81,166 pieces of data crawled on the Weibo platform, the study found that the evolution of public opinion is divided into the “initial period” (July 20-21), “outbreak period” (July 22-25), “recurrence period” (July 26-28) and “slow period” (July 29-August 3), and the core topic has shifted from “heavy rain” and “subway” to policy reflection such as “rescue” and “reconstruction”. Sentiment analysis showed that positive sentiment intensity (peak 0.953) was significantly higher than negative (peak -0.947), and social cohesion was present throughout. The experiment shows that the clustering effect is best when the theme dimension K=200 (F-measure=0.5398, Purity=0.812). This study provides data-driven analysis method support for the dynamic monitoring and governance of public opinion on public emergencies.