Research On The Multivariate System Closed-loop Identification Based On Data-driven

Modern thermal power plant, which is large inertia, nonlinear, and strong coupled, is a typical multi-input multi-output (MIMO) object. The Coordinated Control System (CCS) is the most important control system of power plant, and is regarded as one of the most complex control system in the thermal power plant. With the large-scale wind power integrated into the grid, recent grid codes have specified higher requirements of thermal power plants. For the strong coupling and time-varying, the CCS is very difficult to obtain a desired performance with classical PID controller. The advanced control theories may achieve better performance than conventional PID controllers used. However, those advanced control theories usually need the high accurate model to maintain the desired performance. Industrial processes are general nonlinear and multivariable systems, which are not easy to identify the accurate parameters by using conventional identification methods, such as least square identification (LSI).In recent years, subspace identification method (SIM) has a better numerical reliability and a modest computational complexity, by which the system’s state space models are directly identified based on input-output data. SIM is widely used to identify multivariable systems.This paper introduces the basic principles and classical algorithms of subspace identification algorithm, then analyzes a kind of closed-loop subspace identification algorithm based on orthogonal projection (CSOPIM) and a kind of two-stage closed-loop subspace identification algorithm based on SIM and PEM. By comparing the two closed-loop simulation examples, the results show that CSOPIM can accurately identify the state parameters of the controlled object using closed-loop input and output data. In order to further verify the algorithm in the actual system,, CSOPIM is used to identify the thermal multivariable closed-loop systems. The text takes the CCS of a 330MW subcritical unit for example, and the dynamic models are set up in the 90%,80% and 65% conditions based on real-time data analysis and pretreatment. Finally comparing the step response of model with the measured data in the disturbance experiment of power plant, the results show that the identification model can generally reflect the actual dynamic characteristics of unit.