Attentional neural network for the prediction of gas turbine CO and NOx emission

This study aims to predict Carbon Monoxide (CO) and Nitrogen Oxide (NOx) emissions from industrial gas turbines during operation, using various measured variables (e.g., turbine inlet temperature, exhaust temperature). This prediction task can be categorized as a multivariate time series regression. Deep learning architectures have been extensively used in the literature to achieve this, primarily based on Long-Short-Term-Memory (LSTM) and Convolutional Neural Networks (CNN), which were originally developed for Natural Language Processing (NLP) and computer vision. While LSTM and CNNs are commonly used for time series classification and regression, they do not guarantee the encoding of inherent and important properties, such as the auto-regressive nature of the series and the local and global correlations between variables. To address this limitation, a novel model named TapNet integrating LSTM, CNNs and attention mechanism has demonstrated superior performance over traditional time series models, albeit computationally intensive. A recent alternative, Catch22, requires fewer computational resources and is widely employed for feature-based time series classification and regression. This paper conducts a statistical performance comparison between two approaches: an end-to-end neural network architecture based on TapNet and a feature-based approach using Catch22. The study employs a gas turbine public dataset from the UC Irvine Machine Learning Repository.