Research On Modeling Methods Of NARX Model And Hammerstein Model For Decomposing Furnace Temperature Processes

The mainstream development of modern cement industry is the new type dry process cement production, one of whose core techniques is pre-decomposing technology. The main equipment to realize pre-decomposing technology is decomposing furnace, whose functions are burning fuel, transferring heat between gas and solid and decomposing carbonate. The stability of decomposing furnace has an important effect on the decomposing rate of raw material and the stability of the final production quality and yield. And decomposition furnace temperature is a key technical parameter of stable running condition of decomposing furnace. However, decomposing furnace, rotary kiln and preheater are directly linked, the mechanism of combustion, heat transfer, physical and chemical reaction in the decomposition furnace is complicated, interference factors are many and working condition is easy to change, it follows that the decomposing furnace temperature process has some characteristics, such as nonlinear, strong coupling indeterminacy and other characteristics, so it is very difficult to establish the accurate dynamic mechanism model and data-driven model. Therefore, researching hard on modeling methods of the decomposing furnace temperature process and modeling more suitable for control and simulation have a very important significance for the designing automation of decomposing furnace temperature.This paper is supported by "large rotary kiln intelligent control system" of the national high technology research and development program of China (863). According to big simulation error, poor performance and low efficiency of the existing model for control and simulation applications, this thesis proposes two kinds of modeling methods which can improve the model performance, and use the actual data of decomposing furnace temperature to build nonlinear dynamic model. This main work of this paper is as follows:(1) Aiming at the big simulation error and broad model structure to describe nonlinear of NARX (Nonlinear AutoRegressive with eXogenous Input), this paper proposes a structure controlled NARX model based on multi-step prediction error criterion, the model structure is composed of two parallel modules which make model structure is controlled, a module describes the dynamic nonlinear relationship between output and inputs,and another module describe the linear dynamic relationship between output and inputs. The paper put forward the multi-step prediction error criterion. So the improved NARX model can study suitable nonlinear of decomposing furnace temperature which makes the simulation error smaller. Extreme learning machine (ELM) possesses the advantages of less parameters and learnning faster, so it can be used as the NARX model parametric form. Grid search and cross validation method is used to solve the model order. Model parameters are estimated using the sequential quadratic programming (SQP) method. Finally the structure controlled NARX model based on multi-step prediction error criterion is established. This paper puts forward the structure controlled NARX model based on multi-step prediction error criterion for decomposing furnace temperature, which is chosen to compare with general NARX model based on multi-step prediction criterion and the simulation results show that the model can reduce the simulation error in a certain extent.(2) Existing Hammerstein-OE model for decomposing furnace temperature has small simulation error, but it takes a lot of time to solve. This paper proposes a model to improve the efficiency of Hammerstein-OE model which is called Hammerstein-OBF model based on the regularized ELM:using the structure of Hammerstein model, ELM model is used as the static nonlinear module of Hammerstein model; In addition, parameters of orthonormal basis filter (OBF) is fewer, and the identification algorithm is simple and it is more suitable for the control and simulation, so it is applied to the linear dynamic module of Hammerstein model. In order to further improve the generalization performance of model, the regularization method, group lasso is used to sparse hidden layer nodes of ELM, and get simpler model, Hammerstein-OBF model based on the regularized ELM is established. Compared with Hammerstein-OE model, simulation results show that this method is able to ensure that the simulation error is small and improve the efficiency of calculation, and it is a more efficient way.(3) Structure controlled NARX model based on multi-step prediction error criterion and the Hammerstein-OBF model based on the regularized ELM are compared on the modeling improvemen, generalization performance and the model complexity The results show that structures controlled NARX model based on multi-step prediction criterion is mainly improved NARX model structure, and the multi-step prediction error criterion is applied to select model parameters; Hammerstein-OBF based on the regularized ELM is mainly to improve the modeling efficiency of Hammerstein-OE model; The complexity of Hammerstein-OBF the model is higher; The generalization performance of NARX model is better.