Abstract: Abstract: The fire risk prediction model based on recurrent neural networks was introduced. The model extracts multidimensional features from historical fire alarm data, unit and building basic information, fire facilities situation, as well as inspection and hidden hazards records, and perform the deep learning and model training. The model has been applied in Mianyang, Sichuan, and predicts the fire risk probability of 41 thousand units in Mianyang in the future 90 days. According to the prediction probability, the rules for the selection of "double random, one public" units are optimized, and guide the supervisory staffs to the units with high fire risk. The results showed that, the model has effectively improved the accuracy of daily fire supervision.

Key words: Key words: fire risk prediction, recurrent neural network, precise regulation, fire management