Predictive Modelling of Gas Turbine Emissions based on Generalized Regression Neural Network Method Approach

Gas turbines are essential in various industries but are sources of pollutants such as NO, CO, SO₂, and particulate matter. For several process sectors, Predictive Emissions Monitoring Systems (PEMS) have shown to be a practical substitute for CEMS. By integrating PEMS, the emission prediction technique could be streamlined and a more cost-effective solution that complies to regulations. This study employs the General Regression Neural Network (GRNN) method to predict gas turbine emissions, utilizing a setup with 16 input parameters. The network architecture includes a hidden layer with 32 neurons using ReLU activation and a single output neuron with linear activation, trained over 100 epochs with the Adam optimization algorithm. This configuration aims to improve predictive accuracy and reliability by leveraging GRNN's robust pattern recognition capabilities and the efficient optimization of Adam's method. The model demonstrates excellent performance for predicting CO₂, O₂, SO₂, and NO, based on metrics such as MSE, MAE, MAPE, and R², indicating accurate predictions and a good model fit. However, carbon monoxide (CO) predictions show higher error metrics (MSE and MAE), suggesting greater variability in the predictions.