Objectives: This study aimed to develop a multidimensional analytical model for spatiotemporal trends and risk food categories， with the objective of achieving efficient early warning of food safety risks and providing a scientific basis and decision support for the intelligent and precise management of supervision. Methods: Based on the food sampling data from Jiangsu Market Supervision and Administration Bureau from 2019 to 2024， a spatio-temporal Transformer based multi-class prediction model was proposed. This model deeply explored the evolution patterns of food safety incidents in both temporal and spatial dimensions and established intrinsic associations with risk categories. To mitigate the long-tail distribution problem caused by disparities in data volume across multiple categories， a weighted mean squared error loss function was introduced to optimize model training， thereby enhancing the sensitivity of the model to tail categories. Results: The accuracy of the model was comprehensively evaluated using four high-risk category classification metrics and three all-category risk classification metrics. It not only achieved prediction accuracy comparable to statistical methods in regions where risks had occurred， but also efficiently predicted all-category risks across the entire spatiotemporal scope. Conclusions: The spatio-temporal Transformer multi-class prediction model provides a novel and effective tool for food safety regulatory authorities， enabling the optimization of random inspection strategies and enhancing regulatory efficiency.