Research On Clustering Algorithm And Modeling Method Of Boiler Combustion System

As a complex large-scale energy conversion equipment in the energy industry,the boiler has a complicated and variable combustion process.Due to the tedious nonlinear relationship between the parameters,it is difficult to establish an accurate model for it by analyzing its combustion mechanism.How to mine the useful information in the historical data of the boiler,and then provide the optimal operation guidance for the boiler,is an important idea for the optimization guidance of the heating boiler in the era of industrial big data.The idea of machine learning is applied to the boiler combustion system in this paper.The characteristics of the boiler combustion system operating conditions are studied and the density peak clustering algorithm is improved for the division of boiler operating conditions.A model combining classification network and least square support vector machine is established.Taking into account the characteristics changing in the operating conditions of the boiler combustion system,a model combining classification network and long-short-term memory network was established,and the point prediction was extended to time series interval prediction.The main work and innovations of this article are as follows:1.For the problem that the density peak clustering algorithm’s assignment process is likely to trigger the domino effect and the performance is very sensitive to its parameter cutoff distance dc,a local reachability density peaks clustering algorithm(LRDPC)is proposed,which can not only automatically determine the number of clustering centers,but also improves the robustness and error tolerance of the algorithm.The local reachability density used in LRDPC takes into account the difference in the distribution of data objects in the local space where it is located,which is more reasonable in geometric meaning.Besides,the cut-off distance dc was deleted to enhance robustness of the algorithm.In addition,a new allocation strategy is developed to eliminate the domino effect.Experimental results confirm that the algorithm is feasible and effective,and has better performance on real-world datasets.2.Aiming at the problems of high complexity and insufficient generalization ability of boiler combustion system models,a DWLS-SVM model combined with classification neural network was established to improve the model prediction ability.Due to the features of the boiler’s original operating data are interdependent and redundant,feature screening is used to simplify the model structure.LRDPC is used to automatically determine the number of operating conditions and clustering the dataset.The kernel function parameter σ and the regularization parameter C in the model are very important to the performance of the model,which was optimized by genetic algorithm to achieve the optimal effect.The results show that the sub-work condition modeling method can improve the prediction accuracy.3.The combustion of the boiler is a dynamic process,and the combustion state of each time node must be affected by the combustion state of the previous time node.Aiming at the problem that the traditional modeling method only uses the current operating parameters and there is insufficient utilization of effective information,the LSTM network is proposed to establish a time series model.Taking advantage of the continuous characteristics of boiler operating parameters in the time domain after clustering,a sub-work condition LSTM model combining multi classification neural network was established.Compared with the traditional point prediction model,the relative error of the time series interval prediction model on the test set is reduced by 0.0187% on average,and the root mean square error is reduced by 0.0405 on average.