Dynamic prediction of SO2 emission based on hybrid modeling method for coal-fired circulating fluidized bed

Pollutant prediction for coal-fired circulating fluidized bed units is crucial for ultra-low emission optimization. Accurate prediction models can assist in the control optimization of the unit. Mechanism models are limited by the determination of parameters, coefficients, and fitting functions in the model and require a large amount of operational and unit design data in practical applications. With the development of deep learning, more and more deep learning models are used in parameter prediction. These models suffer from insufficient prediction accuracy when performing parameter prediction tasks due to the lack of a priori knowledge of the mechanism process. This paper analyzed the relationship between the differential equation model under the first-order Taylor expansion and the single-layer Gated Recurrent Unit neural network model. According to the analysis results, this paper proposed a mixed prediction model of SO2 concentration. The ablation study demonstrated the validity of the predictive model structure. The operation datasets of two actual units were used for verification. In terms of MAE indicators, the results of the proposed model on the two data sets are 124.5669 mg/Nm(3) and 178.0473 mg/Nm(3). In terms of MAPE indicators, the results of the proposed model on the two data sets are 5.85% and 14.07%.

Automated summary

Pollutant prediction for coal-fired circulating fluidized bed units is essential for optimizing ultra-low emission. Mechanism models with the determination of parameters, coefficients, and fitting functions have limitations and require a large amount of operational and unit design data. Deep learning models have been used, but they lack a priori knowledge of the mechanism process, resulting in insufficient prediction accuracy.