An enhanced multilayer Res-Informer with Savitzky-Golay filter for predicting mixed CO and NOx emissions in gas turbines

Abstract

Gas turbines emit large amounts of carbon monoxide (CO) and nitrogen oxides (NO_x) when working, and the emission of CO and NO_x poses serious harm to human health and environment. Therefore, accurately predicting CO and NO_x emissions from gas turbines is of great significance. Traditional machine learning algorithms have significant drawbacks in handling long time series data. They typically require complex feature engineering to manage time dependencies, the modeling process is cumbersome and time-consuming, and they are limited in capturing nonlinear features and handling high-dimensional data, as well as effectively dealing with noise and non-stationarity in data. To address these issues, this study proposes an enhanced SGM-ResInformer. This method combines the characteristics of Density-Based Spatial Clustering of Applications with Noise (DBSCAN), Savitzky-Golay filter, multilayer residual network (M-ResNet), and an improved Informer. Data denoising is performed using DBSCAN and Savitzky-Golay filters, M-ResNet enhances the extraction of complex features, better capturing nonlinear relationships in the data, and in the Informer, the original simple MaxPool1d layer in the self-attention distillation layer is replaced with a learnable convolutional layer for attention distillation operations. Experimental results show that compared to the traditional Informer model, the mean square error (MSE) of SGM-ResInformer is reduced by 44.26 %, indicating a significant performance improvement. Compared with other advanced algorithms like Autoformer, SG-informer, Transformer, and LSTM, SGM-ResInformer also shows varying degrees of improvement. Overall, the model excels in improving prediction accuracy, stability, adaptability, and generalization ability, making it particularly suitable for multi-step prediction tasks of complex time series data. These advantages make the model significantly outstanding in the task of predicting emissions from gas turbines.