The software structure， data flow and function modules of a new unit on-line operation optimization system are described. A new conceptualization of unit optimum operation mode is expatiated to define a constrained optimization problem. A steady state judgment method based on tendency distillation is used to eliminate the unstable operation mode in the unit original operation mode database. All unit stable operation modes are divided into various clusters by K-means analysis. The optimum operation mode in each cluster is obtained with the unit net coal consumption rate as an assessment standard. An artificial neural network model for unit optimum operation mode is developed and verified. An example is given to demonstrate the effectiveness of the model. The application results show that the model with better predicting performance and higher calculation speed can provide operators with an on-line unit optimum operation mode.