Research On Data-driven Modeling Of Thermal System Based On Gaussian Process Regression

With the rapid development of computer,sensor network,data storage technology and its widespread used in large-scale thermal power units,massive historical operating data can be saved.Because historical operating data is the most intuitive reflection of unit operating conditions,and with the development of power station informatization,data acquisition has become very easy and inexpensive,which provides a good foundation for developing data-driven models.However,it is often not easy to build a model with excellent performance.How to select and process the modeling data and how to select the appropriate modeling scheme will have an impact on the final effect of the model.Because of this,the data-driven modeling method has become a research hotspot,and has been continuously concerned by researchers.In view of some problems in data-driven modeling of thermal process,this paper focuses on the modeling method based on historical operating data.The following research work has been carried out around data preprocessing,steady state detection,construction of static and dynamic models,and multi-model modeling methods:(1)Aiming at the problems of abnormal data and missing data in the historical operating data,the detection and correction methods of historical data abnormal values were studied.For outlier detection,an outlier detection method based on empirical wavelet transform and local outlier factor is proposed.Firstly,the empirical wavelet transform is used to extract and remove the running trend of the data.Then the local outliers are obtained from the trend-removed sequence.Finally,the box plot is used to adaptively identify the outliers.Aiming at the problem of outliers and missing data in historical data,the method based on sliding window method and Nadaraya-Watson regression is adopted to correct and complete the relevant data.Taking the load data and total air volume data of a 1000MW thermal power unit as examples,the effectiveness of the proposed method is verified respectively.(2)Aiming at the phenomenon that dynamic data and steady-state data alternately appear in historical operating data,in order to distinguish different data,a steady-state detection method of thermal process based on the combination of signal decomposition and statistical inspection method is proposed.Firstly,the empirical wavelet transform is used to decompose the original signal to obtain the operation trend and oscillation information of the original signal,and then R-statistic test method in statistical theory is used to detect the steady-state data of the thermal process.The simulation data and the data of a 1000MW thermal power unit coordinated control system are taken as examples to prove the effectiveness of the method.(3)Aiming at the characteristics of large amount of data,high attribute dimensions and large data repeatability in thermal process steady-state data,a Gaussian process regression modeling method based on sample sparsity and feature variable selection is proposed.Firstly,the sample selection method based on the combination of data similarity and information entropy is used to sparse the steady-state data.Secondly,the method based on Pearson correlation coefficient and least angle regression is used for data fusion and feature selection of modeling feature variables.Finally,the system model is established by Gaussian process regression.A static model of oxygen content in flue gas of a 1000MW thermal power unit is established by using historical operating data.Compared with other methods,it is proved that the model established by this method has smaller model errors and higher static accuracy,and can achieve good prediction results.(4)Aiming at the characteristics of strong non-linearity between variables in the dynamic process,complex and changeable processes,and time delay between input and output variables,a method of dynamic Gaussian process regression modeling based on combined kernel function and delay estimation is proposed.Firstly,gray relational analysis is used to estimate the time delay parameters of variables,and the data set is reconstructed according to the estimated values.Secondly,according to the closure property of kernel function,a local kernel function and a global kernel function are combined to form a new kernel function,and the combined kernel function dynamic Gaussian process regression model is constructed.Finally,the dynamic model of oxygen content in flue gas of a 1000MW thermal power unit is established by using historical operating data to verify the effectiveness of the proposed method.The experimental results show that the dynamic model established by this method has the characteristics of high accuracy and strong generalization ability,which can meet the actual needs.(5)In view of the existence of multiple working conditions and wide range of working conditions in the thermal process,the use of a single model to describe the characteristics of the production process is likely to cause problems such as over-fitting training and poor model generalization ability.A multi-model dynamic Gaussian process regression modeling method based on adaptive fuzzy clustering method is proposed.Using the proposed adaptive fuzzy clustering method to divide the data set and build its dynamic Gaussian process regression model according to the divided sub-data sets.Then the final output model is obtained by Bayesian fusion method based on the prediction variance.Combined with the historical operating data of a 1000MW thermal power unit,the oxygen content model of boiler flue gas is established.Compared with the single global dynamic Gaussian process regression model and the multi model dynamic Gaussian process regression model using other synthesis strategies,the results show that the proposed method has good prediction accuracy and generalization ability of the model.