Research And Application Of Data Driven Modeling Method For Large Thermal Power Unit Control System

In recent years,the market of thermal power units upgrading in China has come to an outbreak period.Establishing an accurate model of the control system is a key point for the control system upgrading which is an important part of the upgrading of the thermal power unit.Therefore,it is very important to study the modeling method of thermal power unit.Based on the characteristics of thermal power unit historical data and the problems in the transfer function modeling method driven by historical data,data driven transfer function modeling methods are studied intensively in this dissertation.Main contents are given as follows:(1)By analyzing the shortcomings of the teaching-learning-based optimization algorithm,an improved adaptive teaching-learning-based optimization algorithm is proposed.The algorithm introduces an adaptive teaching factor and a learning process with teaching signal into the optimization process.A number of different types of standard test functions are applied to test the performance of the algorithm,and the result shows the superiority of the algorithm.(2)For some thermal power units control systems,after the systems entered a relatively stable state,the inputs and outputs are still fluctuating in a certain range.When modeling for this kinds system based on the conventional historical data driven transfer function modeling method,generally,the average value of the relatively stable data is selected as the steady-state component of the system.This selection method is prone to cause the selection error and make the identification result inaccurate.In order to solve this problem,a historical data driven modeling method based on initial steady-state optimization is proposed.The steady-state component value is taken as a dimension of optimization variables.Before the modeling,the data are processed based on the steady-state component given by the optimization algorithm.A negative pressure control system of a thermal power unit is modeled in this method.And the result shows the effectiveness.(3)For thermal power units,most of the controlled plants are self-regulating plants.Besides,some control systems are in an unstable state most of the time.The steady-state data are difficult to obtain.In this dissertation,a dynamic historical data driven modeling method is proposed which is based on initial states optimization and steady-state component estimation.The method is based on the assumption that the system input is fixed at a certain time,then the system output will eventually stabilize at a certain value.The fixed input value and the final steady-state value of the imaginary output are used as the benchmark for eliminating the steady component of the system.The initial states of the dynamic data and the final steady-state value of the system imaginary output are taken as dimensions of the optimization variables.The methods above provide the acquisition approach for initial states and steady-state component.It ensures the implementation of the modeling method.(4)Considering that it is difficult to determine the optimal scopes of the initial states in(3),the state observer is introduced into the modeling method.In this dissertation,a dynamic historical data driven modeling method is proposed which is based on state observer and steady-state component estimation.Models of thermal power units control systems usually can be expressed by several typical transfer function model structures.In these typical model structures,the ranges of parameters are easy to settle.The state observer and typical model structures are applied to observe the states of the dynamic modeling data,and the states are regarded as the new initial states of the system.Then the system is modeled.This method is a good solution to solve the problem that the ranges of the initial states are difficult to be given.The high temperature superheater system and the oxygen content control system of a thermal power unit are modeled.The result shows the effectiveness of the method.(5)When the unknown disturbance data of the system is included in the modeling data,the output of the system contains the control input component and the disturbance input component.It is impossible to establish the accurate model of the system by the conventional method.In this dissertation,an anti-disturbance historical data driven modeling method is proposed.By analyzing the process that the unknown disturbance act on the system,a steady-state component elimination method is proposed which is based on terminal steady-state value.And the anti-disturbance principle of the data processing method is derived.A fast sieving algorithm of steady-state data is proposed which is based on sliding window.It solves the problem of steady-state data sieving.The initial states of the system are acquired by state observation.A coordinated control system of a thermal power unit is modeled.The result shows the feasibility of the method.(6)In order to improve the anti-noise performance of the Luenberger state observer system in(5),the poles of the observation system are placed on the left side of the s plane near the imaginary axis.This method reduces the observation speed of observer and requires more dynamic data to observe the states of the system.To solve the problem above,the Kalman filtering algorithm with better anti-noise performance is introduced to the states acquisition.Considering the time-delay and nonlinearity characteristics of the thermal power unit system,delay time is regarded as a dimension of the optimization variables.And the system inputs are moved parallel to the right,according to the optimal delay time.The idea of piecewise linearization and linear parameter time-varying is introduced into nonlinear system modeling.Finally,an overall performance modeling method is proposed which is based on time-delay estimation and Kalman state tracking.The modeling of the high temperature superheater system of a thermal power unit shows the effectiveness of the method.