Abstract: The features of coke oven flue temperature are analyzed,and a highly accurate soft-sensing model based on error forecasting is established to measure the coke oven flue temperature on-line.In order to linearly fit the relationship between the flue temperature and the regenerator-top temperature,three linear regression models with one variable,two variables and twelve variables are built respectively,and then the features of the three linear regression models are analyzed and compared.The errors produced by the twelve-variable linear regression model,the most accurate one,are fitted and predicted in a multistep way by the Elman neural network model which is based on the temporal difference method.The expert experience is used to integrate the outputs from the linear regression models and the temporal-difference-based Elman neural network model.At last,the more accurate soft-sensing value of the coke oven flue temperature is obtained with these models.The practical application results prove the validity of the presented soft-sensing model.