Research On Key Technology And Method Of Soft Sensor For Thermal Parameters In Power Plant

The energy utilization efficiency of thermal power generation is relatively low,one of the main influence factors is that lots of important technical parameters and economic parameters are difficult to be real-time measured online,such as oxygen content in flue gas,carbon content in fly ash,ball mill load and other thermal parameters related to boiler efficiency closely.Soft sensor technology is one of the effective ways to solve the problem of measuring these parameters,utilizing some parameters liable to be measured through on-line analysis to estimate these variables unable or difficult to be measured.The following researches are expanded focusing on several key technical problems of soft sensor for thermal parameters:1.This paper comprehensively analyses the characteristics of monitoring objects and thermal parameters in power plant as well as discusses main factors which affect the accuracy of soft sensor for thermal parameters.The results of the discussion were summarized as follows:data preprocessing,auxiliary variables selection,modeling algorithm and model structure.2.Two key technical problems are put forward-data preprocessing and auxiliary variables selection.This paper researches on the methods of improving the accuracy and feasibility of soft sensor for thermal parameters in power plant:adequately considering the error analysis and processing of field data in one aspect of data preprocessing;introducing grey theory to make optimal selection in order to improve the accuracy and feasibility of the models in another aspect of auxiliary variables selection.3.Modeling algorithm-supports vector machine-and its own parameters have a great impact on modeling accuracy.This paper presents a modeling method of parameter self-adaptive support vector machine based on training data which can reduce the impact of human factors and uncertainty of the accuracy in the process of parameter selection.The method has been applied to soft sensor modeling of oxygen content in flue gas using actual field data and its effectiveness is proved.4.Combining the sequential minimal optimization(SMO) algorithm that has the characteristics of fast convergence with particle swarm optimization(PSO) algorithm,the modeling method of soft sensor based on PSO-SMO algorithm is proposed in this paper.The new model gets the algorithm parameters used as initial values from training data,using double-layer optimization method to get further optimization,thus improves modeling accuracy and convergence speed of the model. 5.According to the real-time requirements of thermal parameter monitoring in power plant,the modeling algorithm mentioned above is realized by programming under the conditions of typical hardware configuration of a distributed control unit in DCS and in the environment of real-time operating system of QNX.Modeling with field data,actual results meet the requirements of the industrial field,with great engineering practice significance.6.To adapt better to the changes of field conditions and dynamic characteristics of measure object,this paper studies on the structure of soft sensor model and puts forward the modeling method of multi-model dynamic soft sensor.The method has been applied to soft sensor modeling of oxygen content in flue gas and carbon content in fly ash based on M-SMO with historical data under multiple load conditions in power plant.The result indicates that multi-model modeling method can meet the requirements of accuracy under variable conditions in the thermal process.