Research And Application Of Partial Least Squares Modeling Methods Based On Historical Data

Nowadays, the rapid development of information technology in power plant has provided a convenient platform for the study of data based operation optimization. In which, the modeling of complex thermodynamic system based on the huge amounts of data in the power plant real time/historical database is gradually becoming one of the hottest topics. However, the disadvantages of historical data seriously obstacle the development of modeling methods, such as variable multicollinearity, nonlinearity, non-uniformity distribution of working conditions and so on. To solve these issues, the dissertation studies thermal process modeling methods based on partial least squares projection to latent structures, which solves the problems above in a better way. The main contributions of this dissertation can be summarized as followings:1. The characteristics of the power plant historical data is analyzed, three stages of historical data modeling is summarized, namely, data preparation, modeling process and model validation. Some common pretreatments, modeling theories and model validation methods are introduced, and the difference between fitting precision and prediction precision is elaborated.2. The history of PLS development and current research situation are reviewed, the extraction process of PLS which solves multicollinearity is described, and the method of which determines the number of PLS extracted components by cross validation, some auxiliary analytical methods are introduced. Finally, PLS nonlinear modeling methods are summarized.3. With regard to the characteristic of uneven data distribution, three principles sample selection in historical data modeling are put forward and then method of modeling sample selection is also proposed. After the analysis of several common experimental design methods, the uniform design is determined as the principle of modeling sample selection. Finally, the significance of sample uniformity for improving prediction precision is verified through simulation.4. To solve the problem which no data can be required due to multicollinearity, method based on PLS transform and its improvement based on orthogonal signal correction are proposed, and both of their validity are analyzed through simulation.5. As an example of reheat steam temperature system in thermal process, PLS modeling method based on historical data is proposed. Starting from the energy balance principle, method with expected reheat enthalpy rise (which represents the heat-absorbing capacity of unit flow steam) as the dependent variable is put forward and its influencing factors are analyzed. Some variables which cannot be measured but really play a key role in the field are also constructed. The results show that:First, the model established with expected reheat enthalpy rise can effectively reduce the nonlinearity component of reheat steam temperature; Second, modeling sample selection based on uniform design principle can effectively improve the predictive precision of model.