The accurate prediction of the development height of the "two belts" in ji mine is of great significance for the prevention and control of mine water hazards and safe and efficient production. To construct a prediction model for the development height of the two zones of the mine roof in the western region that is threatened by the water hazard of the thick sandstone of the Luohe Formation, this paper takes a certain mine that is severely threatened by this kind of water hazard as the research object, and takes the measured development height of the two zones of the mine as the dependent variable, and six factors such as the thickness of the coal seam, the thickness of the Luohe Formation strata, and the ratio of sand, mud and rock as independent variables. A regression prediction model for the development height of two belts in mines based on iTransformer Long-short term memory(LSTM) is proposed. This model first extracts the dependency relationships and related features among various variables through the iTransformer model, and then inputs them into the subsequent LSTM model. Complete the multiple regression of the development height of the two zones, and on the basis of this model, introduce the QR quantile model to complete the prediction of the probability interval of the development of the two zones in the mine based on the QR-itransformer -LSTM model. The results of model training, validation and testing show that the mean absolute error () of the training set, validation set and test set of the model ranges from 14.79 to 15.38 m, and the mean absolute percentage error () ranges from 13.03% to 39.26%. The root mean square error () ranges from 18.89 to 23.29 m, and the coefficient of determination () ranges from 95.47% to 97.29%. The prediction results have been significantly improved compared with other models. After introducing the QR quantile model based on the iTransformer-LSTM model, the prediction interval coverage probability(PICP) of the 0.95 and 0.8 quantiles is both 1. It indicates that the prediction results are all reliable at a confidence level of 0.8 and above. The research finds that the QR-iTransformer-LSTM model not only has the superior regression prediction performance of the iTransformer-LSTM model, but also surpasses the point estimation prediction mode of the original model. The final prediction result given is a probability interval. The model has a high degree of credibility and can effectively guide production. The research results can provide a new method for predicting the development height of the two zones in mines with similar geological conditions within the mining area, and have certain practical significance for the water prevention and control work of the mines.