Urban fires occur frequently, posing significant threats to public safety in modern cities. Although previous studies have attempted to predict fire risks using regression analysis, traditional regression models are limited in handling complex nonlinear relationships among variables and accurately assessing the weight of each indicator. Additionally, at the level of prefecture-level cities and above in China, a comprehensive fire risk prediction indicator system has not yet been established. This study aims to enhance the accuracy and practicality of urban fire risk prediction. Based on multi-source data from prefecture-level and above cities in China, this research integrates fire risk, social economy, natural meteorology and spatial location to construct a fire risk prediction indicator system from three dimensions: hazard, vulnerability, and exposure. Five machine learning algorithms were employed for modeling and comparison to quantify fire risks and explore their spatiotemporal distribution. The findings reveal that XGBoost performs the best in predicting fire frequency and direct losses, demonstrating strong generalization and reliability. Further spatiotemporal analysis indicates that fire risks fluctuate upward over time, particularly in years with lower humidity, higher sunlight and high wind speed exposure. Spatially, coastal economically developed cities exhibit higher fire-related economic losses, while fire frequency is more concentrated in the central and eastern regions. Conversely, fire risks are relatively lower in the western and northeastern regions.