Application of hybrid neural network structures for modeling and control of combustion process

Abstract

The scientific work is based on the need to develop a combustion process control system that will optimize boiler operation based on information and measurements, as well as take into account innovative methods for assessing process quality. Modern methods of obtaining information about combustion quality (CO and NOx emissions) in combination with control methods make it possible to reduce harmful gas emissions into the atmosphere and efficiently use fuel associated with renewable energy sources. The dynamics of the combustion process of coal dust and biomass are complex; therefore, three selected deep neural network structures were considered for the study: MLP, simple recurrent network, and LSTM (Long Short Term Memory) cells. The study proposes a new hybrid model based on data processing, which uses selected decomposition methods to divide the total parameters of the combustion process into sublayers. In this work, two MRAC systems were developed and compared. The study considered direct forecasting with a five-step horizon. According to the analysis results, the best results for modeling the time series of the combustion process were obtained using the hybrid EWT-LSTM-RELM-IEWT method. The results obtained on the laboratory bench made it possible to develop robust control using hybrid neural networks.