This study investigated the use of Artificial Neural Network (ANN) and Genetic Algorithm (GA) for prediction of NOx emission from grate-kiln iron ore pelletizing process. Artificial Neural Network is widely used modeling system for predicting and optimizing emissions of air pollutants from complex systems. However, finding the optimum technique for training of the ANN is important in designing the model. This study presents two mains algorithms for training the Multi-Layer Perceptron Artificial Neural Network ANN (MLP-ANN) model, for adjusting its weights and biases. Firstly, the backpropagation training algorithm, which searches for the minimum value of the error function in weight space using gradient descent (minimum of error derivative) to adjust the weight accordingly. While the second algorithm was the Genetic Algorithm, which searches for the solution (set of weights) that will produce the minimum value of the error function by changing randomly that population. The model output showed that, the GA approach of ANN had a lower mean square error (MSE), about 0.0082 lower than the backpropagation training algorithm approach that achieved MSE of about 0.0253. Simulating the ANN with GA approach with new CEMS data for NOx and comparing the results produce a difference of less than 0.2% between the real and predictive data. Finally, the same selected inputs were used with historical Sulphur Dioxide (SO2) data from the same process to generate another network for SO2. Simulating this network with new data and comparing with CEMS data gives MSE of about 0.025.