The Research On Key Index Prediction Model Based On Data-driven In Process Industry

The key indexes of production process plays an extremely important role in the industry production control process.In order to ensure the stability of the whole production process,field operators need to test some key indexes based on their experience.The metamodeling techniques have been viewed as effective technique means to predict these key indexes comapared to the expeience of operators.The metamodeling of key indexes can be individed into mechanism-based modeling and data-driven modeing.The mechanism-based modeling tecnique is that the model is estalished based on physical and chemical mechanism,such as material balance,heat conseration and dynamics.Mechanism-based modeling largely depends on the mechanism of industrial production process.It is diffcult to difficult to guarantee accuracy and reliability of mechanism-based model with the process industry becomes increasingly complicated because of lacking of correct understanding of complex mechanism.The data-driven method completes the prediction task by building a relationship model between the variable inputs and target outputs,which can greatly reduce the cost and difficulty of modeling.Feature selection is the most basic part of data-driven method.The accuracy and efficiency of feature subset search directly determine the upper performance of model prediction.Therefore,a good feature selection method is of great significance for modeling.The wrapper feature selection methods were widely used in the prediction of key indexes in the process industry.However,there exist a common problem that the feature subset search is inefficient.Based on the above analysis,a new wrapper feature selection algorithm is proposed to improve the low search efficiency of feature subset.In addition,on the basis of the new feature selection algorithm,a parameter optimization neural network prediction model algorithm framework based on the new feature selection strategy is proposed.The proposed method is applied to predict the valve opening of reheat steam system in a thermal power plant.The specific research work can be summarized into two parts briefly.(1)The wrapper feature selection methods generally have been faced the problem that is their low feature subset search efficiency.To solve the problem,the thesis proposes an improved partial sequential backwoard floating search mutation combinated with TLBO algorithm for feature selection problem(TLBOw IS).The proposed method is based on an evolutionary algorithm with a partial sequential forwoard floating search mutation for feature selection problem(EAw PS),combining with the advantages of efficient TLBO optimization algorithm and no parameter adjustment.The algorithm can improve the efficiency of feature subset search.The simulation test and analysis of EAw PS and TLBOw IS with seven data sets are carried out to verify the effectiveness of the algorithm.(2)The valve opening degree modeling of the reheat steam system in a thermal power plant have been faced with such problems as unclear mechanism,complicated influencing factors and inaccurate valve opening degree.To solve these problems,the thesis proposes a neural network coupling modeling algorithm framework based on TLBOw IS,which is applied to the valve opening prediction of reheat steam system in a thermal power plant.TLBOw IS provides feature subset for the neural network model.The neural network model uses feature subset to obtain the optimal parameter prediction model through parameter optimization,and provides the evaluation index value for TLBOw IS.By means of this coupling,the two models screen out the features that have the greatest impact on the accuracy of the model,obtain the optimal feature subset,and improve the prediction performance of the model.Two neural network methods,RBFNN and GRNN,were tested and analyzed by using the dataset of reheat steam system in thermal power plant,and the validity of the algorithm framework was verified.Finally,a test platform for value opening prediction software of reheat steam system of a thermal power plant was built,and the offline training of the value opening prediction model and the online real-time prediction functions of the valve opening prediction model were realized.The model prediction performance was greatly improved.