Abstract: To effectively excavate the implicit character of gas emission monitoring data, and to prevent the gas dynamical disaster, based on basic principle of Hilbert-Huang transform （ＨＨＴ） method, the cuckoo search （ＣＳ） and extreme learning machine （ＥＬＭ）, the HHT-CS-ELM dynamic prediction model for gas emission quantity was built. The sample series was decomposed into multiple different frequencies intrinsic mode function （ＩＭＦ） by EMD; the instantaneous frequency of each component was obtained by Hilbert transformation, then divided them into higher frequency and lower frequency; different prediction models were used to predict the IMF; the final prediction results were obtained by superimposing each forecast. This paper took the gas emission monitoring data in a coal of Fenxi Mining Industry as an example to carry out simulation experiment. The results show that: the HHT method can effectively reduce the complexity of the monitoring data, and the minimum relative error is 0.144%, the maximum relative error is 0.388%, the average relative error is 0.281%; this model has higher prediction precision and generalization ability; it can be well applied to non-stationary time series prediction.