Circulating fluidized beds (CFB) are widely used in energy industries for increasing the efficiency and reducing environment pollution. CFB modeling and identification have significant importance for operation optimization. Owing to the nonlinear nature of CFB operation, online CFB modeling and identification are highly desirable so that the model can adjust itself according to the change of CFB operation. In this paper, we develop an online CFB identification method based on diagonal recurrent neural network (DRNN) modeling. This method was applied to a large-scale cold flow CFB at the National Energy Technology Laboratory for prediction of solid circulation rate. The result showed that this method worked excellently.