Research On Modeling Method Of Wet Desulfurization System Based On Data-driven

Thermal power generation is the main power generation method in China.The flue gas produced by coal combustion in thermal power plants is one of the important sources of atmospheric sulfur dioxide（SO₂）pollution.In order to reduce SO₂ emissions,the limestone-gypsum wet flue gas desulfurization system（WFGD）is widely used in thermal power plants.In this paper,the SO₂ emission concentration at the outlet of the desulfurization system is taken as the target variable,and the modeling of the wet desulfurization system is studied based on the data-driven method,in order to provide better decision support for the operation optimization and regulation of the wet desulfurization system.The main research contents are summarized as follows:Aiming at the problem of noise in the historical data of the wet desulfurization system,the Complete Ensemble Empirical Mode Decomposition and wavelet threshold denoising method are used to weaken the influence of the noise in the wet desulfurization data on the prediction results.Firstly,the modal components of the wet desulfurization data are decomposed based on the CEEMDAN method,and the noisy components are analyzed and screened,and wavelet threshold denoising is performed,and the components are reconstructed to obtain the denoised pure wet desulfurization data.The experimental results show that the denoised data is smoother and has a higher signal-to-noise ratio,which lays the foundation for model construction.Aiming at the problem that most of the existing modeling methods cannot fully and accurately extract the hidden features in wet desulfurization data,a wet desulfurization SO2 prediction method based on feature fusion and deep learning is proposed.The convolutional neural network is used to extract the data variable features and complete the feature fusion,and the time features in the data are mined based on the long short-term memory network,and the attention mechanism dynamically weights the hidden state of LSTM to solve the problem of decreased prediction accuracy caused by long sequences.The experimental results show that,compared with other prediction models,the CNN-LSTM-AM model constructed in this paper has higher prediction accuracy.Aiming at the shortage of wet desulfurization data samples,a SO2 emission prediction model suitable for small samples was constructed.The original Northern Goshawk Optimization algorithm is improved to enhance its global search capability,and the parameters of the Extreme Learning Machine are optimized to build the optimal model.Further,Time-series Generative Adversarial Networks is used to enhance the desulfurization data.Experiments show that the enhanced data conforms to the characteristic distribution of the original data,and the performance of the model trained with generated data has been improved.