Research On The Pretreatment Methods Of Power Station Historical Data

The large amount historical data of power plant on-line monitoring database is the most truly reflect of the unit operation condition. Deep excavation of the data can provide the reliable data support for thermal modeling and performance optimization. This paper mainly studies the acquisition of modeling samples in the historical database. The proposed method can avoid using the field test method and has the characteristics of short working period and low cost.The rational variable selection can effectively reduce the complexity of the algorithm and improve the speed of computation. The traditional method selects data through the analysis of mechanism. This requires deep understanding of mechanism. In this paper, using attribute reduction method based on Rough Set to select key variables, and according to the attribute importance to remove redundant variables from original database. Due to only consideration of the correlation among variables, this method has universal applicability. Only the steady-state data can represent the reality of unit actual characteristics. Therefore extraction steady state data from a historical database is necessary. This paper combines adaptive Gauss filter and steady-state detection method to retain mutation data as far as possible while eliminating noise. Detection result has higher distinction degree. Considering the capacity of actual historical database to cover various conditions and contain the complexity of data signal, select samples by using uniform test. If all variables in multivariable system adjust the complete division level, the loss rate of sample selection is higher. So use mixed level uniform design to selection sample data. According to the difference of contribution, assign different levels to main variables and secondary variables. The incomplete block design method is introduced for mixed level uniform design table. Using this method can realize generation the test table as demand. Compare with the existing design manual under the same number of tests, this method has good performance to substitute the test table of existing manual. Through sample selection and the modeling simulation of the history data from a power plant1000MW unit1months running, prove effectiveness of the proposed method. Because of the limitation of historical database, the loss of samples inevitably occurs. Apply the theory of completion incomplete information to solve this problem. Then, verify the validity of the method through generating the database and missing data.Finally, the sample selection method is applied to the flue gas oxygen content modeling data selection. Compare the modeling effect by using sample data selected by the proposed method and the data randomly chosen from historical database. The result shows that the sample data selected can well characterize the historical database, and has the characteristics of simplicity and sample stability. And, this method has no requires for mandatory object, can widely used in a variety of research object sample data selection.