Research On Optimization Of Low NO_x Combustion In Coal-fired Boiler Basing On Data Mining

The 13 th Five-year Plan set clear goals for maintaining economic development in a low speed and developing ecological civilization construction. The energy industry, which not only plays as the original power of economic development but also as one of the main pollution sources, has attracted a widespread attention and much research has been done on controlling air pollution of coal-fired power plants. Combustion optimization basing on operational data of utility boiler plays an important role for its economy property, high efficiency, better utilization of information, and especially for the immense potential of data mining.Aiming at data cleaning on utility boiler, which is basic but particularly important work in data mining process, this paper fulfills a cognitive work on operational data and disposes of the dirty point. Furthermore, an algorithm is proposed to tell steady-state cases from historical operational data of utility boiler. Study shows that the algorithm had pretty good performance in detecting steady-state cases.Statistical analysis is used to deal with data from a 660 MW subcritical pressure boiler after data cleaning. The results indicated that with the raise of loading, nitric oxides(NOx) emissions showed a trend of increase firstly and then decrease. While in the same loading level, methods aimed to strengthen the effect of staged combustion, such as amplify SOFA damp feedback and choose a more suitable spout combination, could decrease NOx. For coal blending, both higher volatile and lower ash both could favor NOx decrease. A multiple linear regression model is established to predicted NOx emissions basing on analysis result above and shows a higher accuracy in comparison with model in other's work.Machine learning is used to work on a 700 MW ultra-supercritical boiler. Data-processing is adopted to get the steady-state point which stands for each steady-state case. A method combining statistical analysis and association rule mining is proposed to deal with the point set and draw a conclusion that lower O2, higher furnace pressure difference, and proper burner angle bias all may favor NOx decrease. Bigger opening angle for upper secondary air and SOFA is related to lower NOx emissions to a certain extent. On the other hand, models basing on supervised learning, such as support vector regression(SVR), random forest(RF) and gradient boosting decision tree(GBDT), are built to predict NOx emissions. Feature selection is offered by RF to improve performance of modes. Comparison on prediction accuracy shows that RF-SVR is the most outstanding one.