Research on the prediction technology of SO₂ concentration in outlet flue gas based on LLE-CNN Transformer LSTM

In response to the problems of redundant highdimensional data features, insufficient local information capture, and difficulty in modeling long-term dependencies in predicting SO ${ }_{2}$ concentration in flue gas from coal-fired power plants, this study proposes a multimodal prediction framework that integrates local linear embedding (LLE), convolutional neural network (CNN), Transformer, and long short-term memory network (LSTM). Using LLE algorithm to perform nonlinear dimensionality reduction on high-dimensional running data and extract low dimensional manifold features to simplify model complexity; By integrating the local feature capture capability of CNN, the global attention mechanism of Transformer, and the time series memory characteristics of LSTM, a hybrid model is constructed to enhance prediction robustness. The experiment is verified based on the actual data of a 550 MW unit desulfurization system. The results show that the MAPE of the fusion model is $\mathbf{9 7. 3 2 \%}$, and the performance of the fusion model is significantly better than that of the single model.