Genetic algorithm-based neural network modeling is studied. A MLP model for predicting NOx emission in a coal-fired power generation plant is trained using genetic algorithms. In order to avoid over-training, two data sets are involved, i.e. one data set is used for searching the weights and bias, the other set is used for validation. The fitness function for the GA based training is the combination of the training error and validation error. The GA-based MLP model has been tested over different periods of operation, showing the merits of this modeling technique.