Research On Soft Sensor And Steady State Optimization Of Combustion System In Power Plant Boiler

With the concerning about energy saving and emission reduction of thermal power industry, the combustion optimization issues are receiving increasing attentions. Some important parameters (such as unburned carbon in fly ash) can not be measured accurately, which requires the use of soft-sensor modeling technology; in addition, the first step to optimize the combustion process is building mathematical models between the objective functions (such as thermal efficiency, emissions of nitrogen oxides) and the operation parameters. Due to the multi-variable and non-linear characteristics of the combustion process, traditional linearly modeling methods can not be qualified, thus it is urgent to study and apply new modeling technology.The paper firstly introduced the least squares support vector machine (LSSVM) regression that is characteristic of small sample and quick learning, and established a soft sensor model of unburned carbon in fly ash; then local learning ideology was applied in the soft sensor modeling, where an improved form of the kernel function was used to identify automatically the model parameters, the ulmimate simulation experiment showed that local LSSVM soft sensor model has higher prediction accuracy compared to the global learning one.Principal component analysis (PCA) can be used to pre-process the modeling data of LSSVM to eliminate linear correlation between variables as well as to simplify the LSSVM model structure, which also constitutes the PCA-LSSVM soft sensor modeling technology. However, outliers that exist in the data would affect the results of PCA and LSSVM regression, thus this paper proposed a robust soft sensor model with robustified PCA and Weighted LSSVM (RPCA-WLSSVM) that robustified both two stages of the PCA and the LSSVM regression. Simulation results showed that the proposed soft sensor model has better prediction accuracy and robustness.Finally, steady-state combustion optimization was discussed, which is based on LSSVM modeling and sequential quadratic programming (SQP) algorithm. A simulation experiment of single-objective optimization regarding unburned carbon in fly ash was carried out and the feasibility of the result was analyzed; in view of the contradiction of raising thermal efficiency and reducing NOx emissions, a multi-objective steady-state optimization program was then proposed: LSSVM be used to establish separately the models of thermal efficiency and NOx emissions concerning the operating parameters; evaluation function be used to establishes hybrid optimization goal function; SQP algorithm be used to calculated the optimal value of adjustable operating parameters. Simulation experiment demonstrated the effectiveness and feasibility of the proposed multi-objective optimization program.