Abstract: The mining disturbance of coal mine is easy to induce the dynamic disaster of rock burst， and its evolution process is accompanied by obvious microseismic response. It is the key to the safe mining of deep coal mine to accurately predict the temporal and spatial evolution law of microseismic risk and real⁃ ize the early warning of disaster. Machine learning can effectively mine the characteristics of microseismic monitoring data，capture the law of time series evolution，and provide technical support for the identifica⁃ tion of rock burst precursors. Aiming at the problems of weak spatial continuity of discrete microseismic scatter points，difficulty in accurately characterizing the evolution characteristics of risk field，and insuffi⁃ cient generalization ability of the model，this paper proposes a time-space strong advance prediction meth⁃ od of microseismic risk field that combines data enhancement and deep learning. This method uses a two�dimensional Gaussian interpolation algorithm，sets a 64×64 grid resolution and a time-series sliding win⁃ dow with a capacity of 5，and reconstructs a discrete microseismic sequence into a continuous microseis⁃ mic risk field image frame of the working face. The data enhancement is completed by 0.9 inter-frame simi⁃ larity threshold，which optimizes the sample quality and improves the convergence efficiency of model training. Based on the ConvLSTM network，the spatial and temporal dependence characteristics of the risk field are mined to realize the dynamic advance prediction of the microseismic risk field. Based on the mea⁃ sured microseismic data of 6439 groups of 7303 working face and 4084 groups of 7304 working face in Zha⁃olou Coal Mine，a comparative test was carried out. The 5×5 sliding window peak extraction method was used to invert the daily microseismic maximum energy，average energy and event frequency index and com⁃ plete the quantitative analysis. The experimental results show that the method can accurately construct and predict the temporal and spatial evolution law of microseismic risk field，with low prediction error and good fitting effect. It shows stable prediction performance and excellent generalization ability in different working faces. The research results can provide new ideas for the dynamic classification and early warning of mine rock burst，and provide technical reference for the prevention and control of similar mine dynamic disas⁃ ters.

Key words: microseismic monitoring, rock burst, space-time prediction, ConvLSTM, data augmenta? tion