Data-Based Research And Application Of Process Model Identification, Control And Optimization In Ethylene Plant

Ethylene plant, as a leader in the petrochemical industry, has more energy consumption and large emitters. How to ensure the safety and stability of equipment operation, improve product quality, save energy, and create more efficiency are the important topics of engineers and managers. Realization of advanced process control in the plant is one of the best solutions to solve these problems. The tests on the plant are usually required for traditional advanced process control, but the great scale devices, the nonlinear and strong coupling of the processes make them impossible because these tests could easily lead to plant instability or even producing safety problems. This is why these technologies are difficult to implement. In addition, these techniques require a high level of technician for daily maintenance; however the shortage of such engineers results in that this advanced control can’t be operated for a long time.The development of information technology makes vast amounts of plant data to be stored. The data have a lot of information about changes in technology and production operation, and this information is needed for advanced process control. How to obtain required data for advanced process control for these data and how to implement advanced control based on these data are the keys to solve the above problems.Based on plant data and combined with modern control theories, a research and an implementation of advanced process control for ethylene plant are carried out. The main works are as follows:1. Study for solving inequality constrained random orthogonal optimization algorithm. Implementation advanced process control in a plant needs accurate process models, and some optimization algorithms are also needed to complete this task. In order to improve the accuracy of search optimization algorithm and to achieve higher efficiency even with higher dimension parameters, a study is done based on NLJ (New Luus-Jaakola algorithm) random search method and orthogonal experimental design method. A stochastic optimization algorithm (ROAS) is proposed to solve problems with inequality constraints.Some groups of initial search vectors are firstly produced, and each search vector is optimized by simplified NLJ algorithm. A global optimized vector is obtained from all of the vectors. Given search radius and based on initial search vectors, the iterative orthogonal experimental analysis of optimization is used to optimize the value of each search vector. The iteration is executed by the following steps:Orthogonal experiment analysis method is applied to search the optimal values. Cluster analysis methods are used from obtained search vectors to produce three level data sets of orthogonal experimental analysis. The orthogonal experiment analysis optimization is carried out with those three level data sets. The global vector is updated with these optimization results. The search vector with maximal objective function value is changed by global optimized vector while other group vectors are updated by the global vector. Finally, narrowed the search radius, a new iteration of the search is started.2. Research problem of continuous-time process model identification problem with colored noise. Noise is existed inevitably in plant data and many of them are colored noise; therefore, it is necessary to identify the noise model while the process model is identified. The colored noise model is defined as an ARMA process. In order to improve the accuracy of identification, a maximum likelihood parameter estimation algorithm with a penalty factor (PML algorithm) is proposed. The method adds a penalty part in terms of the original maximum likelihood parameter estimation algorithm for improving identification accuracy.To achieve rapid and effective identification of the dynamic response model, a hybrid model identification algorithm is proposed. Hybrid models are a kind of Box-Jenkins models consisting of a continuous-time transfer function process model and a discrete ARMA noise model. An instrumental variable method combined with PML algorithm (IV-PML) are used and input output data are dealt with linear pre-filtering for direct identification open-loop system model from sampled data.The measurement noises in closed loop data make the input and output data correlated, so some open-loop identification method is difficult to be applied. A new cycle noise eliminating identification method (DCIV) is proposed to solve this problem. The method, like open loop identification, is a hybrid Box-Jenkins model identification method too. DCIV algorithm calculates the input cycle noise component and the cycle output noise component according to the initial value obtained by IV-PML. Then hybrid Box-Jenkins model identification is carried out with cycle noise free data, which is achieved by subtracting the cycle noise components from the original input and output.3. Investigate issues of identifying automatically process model set and tuning controller parameters based on the model set from plant data. Due to production and safety reasons, the traditional model identification method based testing process can not be implemented in large-scale ethylene plant. An automatic model identification method is proposed based on plant data for solving this issue. Process object has not same parameters at different points of plant load. It is necessary to obtain the process model set under different conditions and tune controller parameters such that the controller has good control performance and robustness.In order to obtain the desired valid data sets for process model identification, a dynamic response data set with tri-state is proposed for choosing the feasible data set from the plant data. Ⅳ-PML or DCIV identification algorithm is applied to identify each response data set. The confidence function of identification is used to decide which feasible data set can be a valid data set for process model identification. The process transfer function model with time lag is identified by ROAS algorithm in terms of valid data sets.The process model set is used to design the PID controller parameters so that the loop can obtain more robust control performance with the optimized parameters. A model set based PID and IMC-PID parameter tuning algorithms are proposed. The ROAS algorithm is used to optimize a set of PID parameters based on all the process models so that the PID loop has a good control performance and robustness for all of process models in the model set.4. Study modeling and optimization of the process based on RBF neural network. To improve the predictive ability of RBF network model, a self-learning dual-model RBF neural network (SODM-RBF) is proposed. The width of the Gaussian kernel optimization based on gradient descent algorithm is proposed. Each hidden layer node center is optimized in terms of these optimized width values.The soft instrument of ethane concentration in ethylene tower and the soft instruments of propane concentration and propylene concentration in Propylene tower were established by SODM-RBF. The optimization control of ethane concentration, propylene concentration and propane concentration are realized.5. Practical application of the above solutions. Based on long-term control engineering work and the problems found from practice, researches of control theory abstracted from these problems are carried out. The research findings have been successfully applied in Practical application, and this is an important feature of this paper. IMC-PID tuning parameters are applied to the ethylene plant. These PID parameters are designed based on identified process model set. The PID loops with these optimized parameters have some good control performance such as fast, no overshoot or small overshoot. Cracking furnace is a complex system with many control loops. Control loops of the furnace have some mutual coupling serious phenomenon between them. The PID loops with optimized IMC-PID parameters ensure each control loop has a fast and smooth operation. A smooth and coordination control for whole cracker furnace are obtained based on fast and smooth single loops operation and all subsystems is considered as a entirety.