Research And Application Of Parallel Hybrid Modeling Method For Thermal Power Units

Improving the economic operation ability of the unit,promoting the verification and application of advanced control algorithms,enhancing the simulation accuracy and training effect of the simulator,and improving the accuracy of fault diagnosis all depend on the system model of the thermal power unit.However,due to the characteristics of large delay,large inertia and strong coupling of the unit,it is very difficult to establish a high-precision mathematical model.Therefore,it is very important to study the modeling theory and method of thermal power units and establish a model that accurately reflects the internal operation mechanism of the system.In the modeling method,the mechanism model has strong explanatory ability and good generalization ability,but it needs a lot of prior knowledge;The data-driven model has strong nonlinear fitting ability and does not need a detailed understanding of the internal mechanism,but relies on massive data;Hybrid modeling combines the advantages of the two models,which makes it have strong nonlinear fitting ability and generalization ability.When the mechanism model is mismatched with the actual system,the data-driven model can compensate the error and improve the overall accuracy of the model.Therefore,the hybrid modeling method is more suitable for establishing the high-precision model of thermal power unit.Field data is the most intuitive reflection of unit operation,and the effectiveness of data directly affects the accuracy of the established model.Because hybrid modeling includes data-driven model,it requires high data quality and needs data preprocessing.The calculation criteria and filling strategy of sample similarity matrix of random forest are improved,and the missing value is filled based on the improved random forest.For the isolated tree,the difference measurement method based on regularized mutual information and the accuracy evaluation method based on pseudo label are proposed.The redundant tree is found by genetic algorithm to optimize the isolated forest,and the abnormal samples are detected based on the improved isolated forest.In this paper,RBF neural network is selected as the data-driven sub model in the hybrid model,and the topology,training process and optimal approximation characteristics of RBF neural network are described.In view of the shortcomings of the general parallel hybrid model,combined with the autoregressive characteristics of process data,this paper proposes a parallel hybrid modeling method with error feedback,analyzes the architecture of the model,gives the model training method,and analyzes the historical information length of feedback based on Spearman correlation coefficient.Taking the condenser of a 1000 MW thermal power unit as the research object,its mechanism model,data-driven model,general parallel hybrid model and parallel hybrid model with error feedback are established and simulated respectively.The method to determine the key parameters of the mechanism model is given,and the structure and parameters of RBF neural network are initialized through cluster analysis.The comparative analysis is carried out from three aspects of model accuracy,generalization performance and convergence performance,which reflects the advantages of parallel hybrid model with error feedback.Aiming at the defects of the simulation support system,the advanced model algorithm container is developed,the algorithm encryption method is given,the interface specification is designed,and the script call scheme based on producer consumer mode is provided.The condenser hybrid model established in this paper is embedded into the simulation support system for configuration and application.