Abstract: For low prediction accuracy of D-S evidence theory used in water inrush prediction of coal floor because of poor robust and fusion effect under the condition of large conflict probability of evidence source, an adaptive robust evidence theory was proposed based on robust evidence theory. A water inrush prediction model of coal floor was established by combining the adaptive robust evidence theory and multiple neural networks. The model was experimentally researched by use of real hydrogeololgy data of a coal mine working face. The experimental result shows that the model has high prediction accuracy and good stability.