Prediction of SO2 emission concentration in industrial flue gas based on deep learning: The ammonia desulfurization system of the Yunnan aluminum carbon plant as the research object

Abstract

The desulfurization efficiency of flue gas desulfurization (FGD) systems is affected by many operation parameters, and predicting SO2 emission concentration through mechanism models requires a significant number of resources, limiting the process optimization of the overall desulfurization efficiency. Taking the desulfurization system of Yunnan aluminum carbon plant as the research object, based on the inlet flue gas, process control parameters, and pollutants of the discharged flue gas, we constructed the NARX (nonlinear auto-regressive model with external inputs) neural network and TCN (temporal convolutional network) models. Predicted the concentration of SO2 in the discharged flue gas in the next moment from the above three types of historical data, optimized the industrial control parameters, and reduced the concentration of the discharged SO2. The results show that both TCN and NARX models have strong nonlinear dynamic descriptive ability, and both are capable of accurately predicting the SO2 concentration in the emitted flue gas. The TCN neural network model can control the mean absolute percentage error and root mean square error of the dataset within 1.46% and 1.86, respectively. The response relationship established by the model has an important role in guiding the regulation of SO2 emission and process operation parameters, such as reducing the flue gas pressure of the absorber tower inlet or increasing the pressure of the concentration section of the absorber tower, which can effectively improve the desulfurization rate.