Real-time prediction of SO2 emission concentration under wide range of variable loads by convolution-LSTM VE-transformer

In this paper, a sulfur dioxide (SO2) emission concentration real-time dynamic prediction model is proposed to achieve the control target of ultra-low SO2 emissions , economical desulfurization system operation at a large-scale coal-fired power plant. First, a vision-expansion self-attention strategy is adopted on the Transformer al-gorithm structure for better time-series information feature extraction. Second, to adapt to the vision-expansion self-attention strategy, convolution and long short-term memory (LSTM) are combined to process input infor-mation position encoding, which can obtain long time-series information absolute position encoding. Then, a Convolution-LSTM VE-Transformer model is constructed by combining the above two proposed parts. The mutual information method and the differential evolution algorithm are combined to calculate correlations between variables and obtain the optimal delay time. Finally, a real-time dynamic prediction model of SO2 is established. In this model, the structure of the input and output variables is determined by the delay time and the influence of coal quality on the SO2 concentration. The quantitative comparison results show the superior ac-curacy and generalization ability of the proposed prediction model to related popular methods, indicating that the model has strong potential for use in further controller design and applications.