Abstract: In order to predict the filling performance of caves in carbonate reservoirs, using the feature that the BP neural network can realize the linear mapping between input and output targets, and based on the detailed characterization of underground river network, ancient landform, ancient drainage system and passageways of underground river, the paper classifies the underground rivers, uses the data of logging, 3D seismic and cores to describe the cave filling degree penetrated by a single well, and predicts the filling degree of the underground river in Lungu oilfield by combining with the correspondence relationship between the cave filling degree and through-well underground river. The dissection of the underground rivers in the study area shows that there are 6 factors which can control the filling degree of the underground river such as the level and type of underground river, cave type, the relationship with the underground river passageways, the relationship with hall-like caves and cave collapse. Using the 6 factors as the input parameters, a prediction model is established for the underground river filling degree. The prediction results show that the relative error is about 10% by using the BP neural network to predict the underground river filling degree. The method can be applied in the evaluation of caved carbonate reservoirs