Research On Flame Combustion State Assessment And Prediction Method Based On Multi-source Heterogeneous Data

Under the background of China’s "3060" carbon target and flexible transformation of thermal power units,steady combustion of boilers which are under low load condition is the key factor limiting thermal power units on reducing the minimum technical output.Therefore,there is an urgent need to establish a set of accurate and timely assessment and prediction methods for boiler combustion status,which plays an important role in safe and economic operation of the unit and reduction of pollutant emissions.Considering that flame image is the most intuitive form to present the combustion state of pulverized coal,and other process parameters such as coal feed and primary air volume also have a certain degree of influence on the combustion state.This paper proposes a quantitative assessment and prediction method of combustion state based on flame image and a prediction method of combustion state based on multisource heterogeneous data,and verifies the rationality and effectiveness of the proposed method using data from a 660 MW coal-fired power plant.The research work and conclusions of this paper mainly include:(1)An unsupervised flame potential feature extraction method and a quantitative assessment method of combustion state based on flame images are proposed.Firstly,a stacked convolutional autoencoder(SCAE)is introduced into the convolutional block attention module(CBAM)to build a flame potential feature extraction network.Next,the flame potential features are further dimensionality reduced,smoothed and normalized,and the combustion state index is defined to assess the combustion degree of pulverized coal in the furnace.Finally,the combustion stability result is obtained by calculating the variance through the sliding window to assess the stability of the flame during combustion.The experimental results show that the extracted potential features of the flame can reflect the key information of the original images,and the proposed combustion state index and the combustion stability result can achieve quantitative assessment of the combustion state.(2)A combustion state prediction method based on flame timing characteristics is proposed.A bi-directional long short-term memory(BiLSTM)neural network model is constructed by considering the flame burning trend to predict the combustion state,taking into account the time series forward and reverse timing information patterns.In order to optimize the model performance,a Bayesian optimization(BO)method is proposed to find the optimization of the model hyperparameters.The experimental results show that the proposed combustion state prediction method has faster convergence,better fitting accuracy and better prediction performance compared with other methods.(3)A combustion state prediction method based on multi-feature fusion and combined model is proposed.Firstly,the flame potential features are extracted from the flame image,and the process parameters are extracted from the DCS using recursive feature elimination method.The two are concatenated as the input features of the combined model.Next,the BiLSTM model and the light gradient boosting machine(LightGBM)model are trained separately.Finally,the two models are combined and the weights of the two models are optimized using particle swarm optimization(PSO)algorithm.The experimental results show that the method of using multi-feature fusion and combined model has good effect on improving the model prediction accuracy.