With the rapid development of social media platforms, the dynamics and complexity of online public opinion have posed serious challenges to disaster emergency management. In this paper, we propose a risk control framework based on deep reinforcement learning, which realizes structured modeling and dynamic control of public opinion evolution by constructing a spatio-temporal decomposition meta-model of public opinion event scenarios, decoupling the events into three types of characteristic elements, namely, location, type of public opinion, and subject of public opinion, combined with knowledge meta-theory, and designing the Weisbuch-Deffuant viewpoint aggregation rule that takes into account the heterogeneity of the individuals and time lag of the environment. In order to realize the structural modeling and dynamic control of opinion evolution. For sensitive text classification, a heterogeneous graph construction method is proposed to integrate text, words and sensitive entities in the public opinion domain, and graph convolutional networks are utilized to enhance semantic association and risk feature extraction. The “2023 Beijing-Tianjin-Hebei Heavy Rainstorm” event is used as an empirical case to analyze the sentiment, influence and stage evolution of public opinion. The experiment shows that the classification accuracy of the risk warning model based on Weisbuch-Deffuant network reaches 99.36%, and the root mean square error (RMSE) is 0.0079, which is 6.46% and 1.97% lower than that of the traditional BP and GA-BP models, respectively. The risk level analysis shows that most of the public opinion events are concentrated in medium risk (C level), which verifies the effectiveness of the model in dynamic prediction and precise warning.
