Fault Detection And Dynamic Modeling Of Utility Boilers Based On Historical Data

As clean and renewable energy such as wind power and photovoltaic holds a growing share of the power grid in China,the power grid puts forward higher requirements on the peak shaving capacity of thermal power generating units to ensure safe and stable operations and consumption of new energy.However,the frequently and rapidly load changing operation and the long-term low-load operation have caused the life of the unit equipment to be severely affected.On the other hand,the operation control of the unit has become more difficult.In view of the above problems,this paper makes full use of the potential knowledge contained in the unit's historical operating data,adopts data mining and machine learning techniques,and realizes data-driven fault detection and dynamic modeling.It is of positive significance for timely and early detect faults in operation,master the dynamic characteristics of the operation,and better control pollutant emissions.The content of this article is as follows:(1)Data is the basis for data-driven applications.The historical operating data of thermal unit is different from the data obtained from well-designed field experiments,with abnormal values,instability,etc.Firstly,this paper introduces the data preprocessing methods,ANN modeling methods and gradient descent training algorithms commonly used in data-driven applications.Then,in light of the operating characteristics of thermal power plants,the problem of the steady-state selection of historical operational data is emphatically explored.(2)In order to solve the problem of steam leakage in the waste heat boiler of a 390 MW gas-steam combined cycle unit,the principal component analysis method was used to re-project the operating data into its principal component subspace and residual subspace,and through the Hotelling's T~2 statistics verification and SPE statistics verification,respectively,to determine if a fault has occurred.The results show that both Hotelling's T~2 statistics and SPE statistics can catch abnormalities in the operation data in time,supporting the next fault diagnosis.At the same time,it was found that the Hotelling's T~2 statistic can more clearly and sensitively react to abnormalities in operation data than SPE statistics.(3)Based on historical operating data,LSTM was used to establish a dynamic model of the NOx concentration at the SCR inlet of a 660 MW coal-fired boiler.The hyperparameters of LSTM model and the learning rate of training algorithm are studied in detail.And the effect of using traditional ANN modeling under the same conditions is compared.The results show that the LSTM has a high prediction accuracy and can handle the load changing operation of thermal unit.However,due to the poor statistical characteristics of the operating data in variable conditions,the prediction accuracy of traditional ANN is seriously deteriorated,and it cannot adapt to the dynamic modeling of the operation process.