The heating values of municipal refuse entering into the incinerator are unstable, which have a great influence on the stable operation of the incineration. By using genetic algorithms to optimize initial weighs and thresholds of BP neural network, a predictive model is established to predict the heating values of municipal refuse. The online operating data are processed based on Garson method and principal component analysis, and then those data are used as input parameters of BP neural network. The measurement and prediction of the heating values of municipal refuse are able to be obtained on-line. The results show that the relative average error for the predicted values is 2.64%. The confidence interval (95%) of the relative average error for the test samples is between -1.75 and 2.59. Therefore, the model has both high accuracy and high confidence level, which is very suitable for engineering applications.