The Research And Application Of Data-based Process Modeling Methods

To ensure the secure and economical operation of the power plant, some important variables are required to be measured accurately and reliably. However, it is very difficult to achieve such measurements only by using hardware-based sensors owning to economic and technical limitation. The data-driven soft sensor has been used as one of the approaches to overcome the problem. Especially the development of informatization of power plants, which makes it easier to obtain the operation data, has provided a favorable research platform for developing data-driven models. The data-driven modeling technique has become one of research focuses in the thermal process field.Commonly there mainly exist two separate sources of data samples for the model development:experimental data samples which are gathered through specially designed experiment and plant data or historical data samples which are captured from historical normal-operation database. The two types of data samples have different characteristics in terms of the distribution, steady condition, uniformity, correlation, samples quantity and etc. Based on the analysis of data characteristics and data preprocess, the modeling methods are studied mainly on the following aspects:(1) Steady state detectionA steady-state detecting method is proposed based on piecewise curve fitting considering the operation characteristics of historical data. Discrete sampling points are transformed to continuous signals utilizing piecewise curve fitting method. Moreover, the first and second order derivative sequences are obtained, and meanwhile, high-order noise is also suppressed. The process trend and steady information are extracted based on threshold criteria. An application is explored using the operation data of feed-water flow system from a600MW utility to validate the effectiveness of the proposed approach.(2) Nonlinear PLS model integrated with inner LSSVM functionConsidering the characteristics of the experiment data, this dissertation presents a new nonlinear partial least squares modeling method, in which outer linear partial least squares (PLS) framework is applied to extract the input and output components and eliminate the correlation. Meanwhile, least squares support vector machines (LSSVM) is deployed to describe the inner relationship between the components. Moreover, the weight-updating procedure is incorporated to enhance the accuracy of prediction. The model validity is examined based on a pH neutralization benchmark process. Then the proposed method is utilized to predict NOx emission using the experiment data of a coal-fired boiler and high accuracy is obtained.(3) LSSVM-based ensemble modeling methodA new LSSVM-based ensemble algorithm is proposed to tackle the problem of historical operation data being concentrated in local regions and of large sample size. Based on traditional fuzzy c-means cluster (FCM) algorithm, a soft cluster (SFCM) method is proposed to divide the initial data into several overlapping subspaces, in which the LSSVM learners are trained respectively. According to the selective ensemble, PLS is deployed as the combiner to obtain the final output with capturing the diversity. The proposed method is applied to develop the NOx model based on the operating data selected from a660MW utility coal-fired boiler, and the comparison results reveal that the time complexity is reduced and the generalization is enhanced.(4) Updating strategy of LSSVM-based ensembleAiming at the time-varying characteristics of industrial process, an updating strategy of LSSVM-based ensemble method is presented. Samples addition and replacement methods are proposed to tackle the extension and transition of the operation region, based on which, incremental LSSVM is employed to accomplish the update of the ensemble model. Meanwhile, the updating time and frequency are also studied. The updating strategy is validated based on the sinc function simulation. The NOx ensemble models with and without update are compared, and the results reveal that the prediction accuracy of the model with update still remains high even if the process characteristics have varied.