Leveraging long short-term memory networks and transfer learning for the soft-measurement of flue gas flowrate from coal-fired boilers

Abstract

The dynamic operation and deep peak-shaving of power-generating units cause significant fluctuations in flue gas flowrate, thus affecting the accuracy of CO₂ emissions measured by continuous emission monitoring systems (CEMS). This study established a long short-term memory network with an attention mechanism (LSTM-AM) for the soft measurement of the flue gas flowrate in real-time. First, flue gas flowrate data and continuous operation parameters over 25 days were sampled from a typical 660 MW coal-fired boiler in China. Then, a carbon balance model was established to verify the data reliability. The LSTM-AM model was trained and testified at the 660 MW coal-fired boiler. Results show that the LSTM-AM model significantly surpassed the pristine LSTM model without attention, the convolutional neural network (CNN) with LSTM, and the static support vector regression (SVR) model in the real-time prediction of flue gas flowrate. Finally, the LSTM-AM model was generalized to a 630 MW coal-fired power unit via transfer learning, which was further demonstrated to outperform the model re-trained from scratch. This work manifests the feasibility of deep learning for the soft measurement of flue gas flowrate, which is promising to solve data-lagging issues when measuring CO₂ emissions from coal-fired power plants.