An empirical model to predict the boiler efficiency and pollutant emissions was developed with artificial neural networks based on the experimental data on a 360MW W-flame coal fired boiler. The temperature of the furnace was selected as an intermediate variable in the hybrid model so that the predictive precision of NOx emissions was enhanced. The predictive precision of the hybrid model was improved compared with the direct model. Three optimal operational objects were proposed in order to minimize the fuel and environmental costs. Based on the neural network model and optimal objects, a genetic algorithm was employed to seek real-time solution every 30 seconds. Optimum manipulated variables such as excess air, primary air and secondary air were obtained under different optimal objects. The above algorithm interconnected with a distributed control system (DCS) formed the supervisory control and achieved real-time coordinated optimization control of utility boilers.