The invention relates to a method for predicting volcanic rock facies, and relates to the technical field of petroleum exploitation reservoir engineering. The method comprises the following steps: (1) forward modeling, (2) seismic attribute extraction, (3) attribute optimization, (4) volcanic rock facies classification, (5) multi-attribute neural network pattern recognition and (6) rock facies prediction, wherein in the step of forward modeling, profile forward modeling is performed on the basis of seismic geologic horizon calibration; in the step of seismic attribute extraction, horizon-oriented seismic attributes are extracted in allusion to a target horizon, and an attribute time window range is determined; in the step of attribute optimization, seismic attributes reflecting volcanic rock distribution are optimized, and preprocessing is performed; in the step of volcanic rock facies classification, category classification is performed on the volcanic rock facies according to a well drilling result; and in the step of rock facies prediction, a plurality of seismic attributes are fused according to neural network parameters acquired by training, and a distribution variation law of the volcanic rock facies in an unknown area is predicted. The prediction method provided by the invention fuses multiple seismic attributes and well drilling information, overcomes the defect that a single seismic attribute only partially reflects the lithological combination difference, and can effectively predict spatial distribution of the volcanic rock facies.