Dynamic modeling of SCR denitration systems in coal-fired power plants based on a bi-directional long short-term memory method

Abstract

Selective catalytic reduction (SCR) denitrification system can effectively reduce NOx emission by controlling ammonia injection. However, the energy structures, load fluctuations, reactor dynamic characteristics and system delay pose great challenges to the precise ammonia injection. To achieve high-precision NOx emissions prediction, a method that combinations dynamic joint mutual information and Bi-LSTM is proposed, where the dynamic joint mutual information theory is used to estimate the reactor dynamic characteristics and system delay. And then, the inputs of the Bi-LSTM are reconstructed according to the estimations. Thus, the Bi-LSTM is established to realize the accurate NOx estimation at the current time and t+3 moment of SCR outlet. Taking a 660MW tangent coal-fired boiler as an example, we establish the Bi-LSTM network by using more than 15,000 sampling data over 11 consecutive days, and predict NOx emissions. Experiments demonstrate that considering the dynamic joint mutual information and reconstructing the inputs, the Bi-LSTM network can greatly improve the prediction accuracy, which provides the basis for the realization of accurate ammonia injection and reduction of NOx emissions.