**2.2 Model predictive control and other control techniques**

Due to the complexity of the process dynamics involved in reactive distillation, conventional control technologies, e.g. PI, PID control, cannot provide satisfactory control performance, while the application of modern control technology requires good process models. A reasonable process model as described by Sneesby et al. [24] contain hundreds of equations for the RD process which is to be controlled. Patternbased predictive control (PPC) is such a method that does not rely on exact process models while providing improved control performance for complex processes over conventional, e.g. PI, control algorithms. Some progress has been made in this direction, like for time delay compensation, Zhao et al. [25] worked for dynamic models that reject disturbances. Since various chemical processes possess time delays and uncertainties, for example, the flowing fluid in a pipe was taken as time delay variable. To represent such system the proposed model works on first order lag dynamics to compensate for uncertainties. Bode plots were also constructed to show that pattern based fuzzy predictive allows a trade-off between robustness and the performance. Seem et al. [26] have proposed a novel predictive scheme by considering a proportional integral controller in which the gain and integral time is calculated automatically and hence they have given such system a name of self-regulating system. However, this scheme was based on the pattern reorganization methodology, but author has asserted that this scheme requires less memory and is more efficient as compared to the conventional techniques. Jang et al. [27] worked on fuzzy predictive control which does not depends on exact process model but considers a pattern predictive control that provides improved performance in both set point tracking and disturbance rejection, shown for the RD process. Local optimum was identified to minimize prediction error and global optimum was then identified through various subsystems. The nonlinear transformation, feature pattern extraction, and PP design was discussed in detail by Tian et al. [28] who have designed a pseudo input– output linear process gain, which needs only a rough and easily obtained knowledge of the steady state characteristics of the process. Author has worked on one-point control strategy i.e. control of bottom purity by considering reflux ratio and reboiler duty as the variables. They also used state estimation approaches for measurement of sample at multiple rates.

Bansal et al. [29] have developed an algorithm for solving Mixed Integer Dynamic Optimization (MIDO) problems. This algorithm is different from other conventional algorithm as they do not depend on the use of a particular primal dynamic optimization method. However, overthe last decades, many versions of Extended Kalman filter (EKF) has coined that deals with the measurements at multiple rates. Patwardhan et al. [30] have worked on output feedback system

for the case of plant mismatch. The EKF for nonlinear systems has been explored by Beccera et al. [31]. The concept involves the use of a time varying linearization of differential algebraic equation in which estimation was performed using a simplified EKF that was integrated with the differential algebraic equation model which accommodate measurements obtained only from the differential states. This technique has serious limitations. This limitation has been overcome by Mandela et al. [32]. They presented formulations for EKF and Unscented Kalman Filter, in which measurements of differential as well as algebraic states were recorded. To achieve more rigorous control of reactive distillation nonlinear system over a wide operating range, various successive linearization based nonlinear predictive control scheme was developed by Huang et al. [33]. They proposed error feedback scheme that introduces integral action in the controller for controlling a multi rate sampled data system. Author has not implemented the conventional linear feedback but adopted a novel variable feedback concept that effectively reduces the noise making the system quite robust for the designer. Akesson et al. [34] have developed a scheme in which control objective was to keep the output close to a specified reference trajectory in such a way that large control signal variations are avoided and possible hard constraints on the state and inputs was satisfied. Main control objective of this work was to minimize the cost. Procedure was adopted by the authors in which controller was trained directly to minimize the cost for a data set, without having to compute the optimal MPC control signals by off-line optimizations.

MPC are classified to various type such as dynamic matrix control (DMC), quadratic dynamic matrix control (QDMC), robust multivariable predictive control technology (RMPCT), generalized predictive control (GPC), and other advanced control techniques and was reviewed in brief by Sharma et al. [35]. They also presented work on comparison of conventional strategies with MPC and neural network predictive control by considering a TAME reactive distillation column for different load changes and proved that NNPC and MPC provide much accurate result as compared to conventional PIDs. **Figure 2** shows the general MPC structure.

## **2.3 Control of reactive distillation using soft sensors/soft controllers**

Design of a soft sensor for a reactive distillation column includes three steps: first, selection of secondary measurement of the process, second, moving data collection and processing this data, and the last, modeling of process based on selected

**Figure 2.** *General MPC structure.*

#### *A Review on AI Control of Reactive Distillation for Various Applications DOI: http://dx.doi.org/10.5772/intechopen.94023*

secondary measurements and processing of data. Generally, the differential algebraic equations (DAE), describes the dynamics of a reactive distillation column. Soft computing techniques are methods in which real practical situation could be dealt in the same way as human deals them i.e. based on intelligence, common sense, reasoning, analogies, etc. Fuzzy logic is the oldest control schemes used not only in process industries but in vast area of other engineering applications also. In view of this, Babusk et al. [36] have first coined up the detailed concept of fuzzy logic in both static and dynamic system and proved fuzzy system as an interactive method, facilitating the active participation of the user in a computer-assisted modeling session. The fuzzy model proposed by Takagi and Sugeno described by fuzzy 'IF-THEN' rules represents local input–output relations of a nonlinear system. Rico et al. [37] applied the fuzzy control technique to control the process output from distillation column inthe desired range for different input disturbances. As an initiation to fuzzy logic, industries as well researchers move toward the field of soft sensing and control using soft computing techniques. These techniques were initially based on local optimization such as given by Pekkanen et al. [38] for a stage by stage specification of reactive distillation. They initiated the control procedure from each column ends i.e., from top as well the bottom while making the design specifications at each stage.

Soft controllers are also known by an alternative name known as intelligent control technique or inferential control as these controllers can estimate and control the process based on past experiences. Use of soft computing approaches take its first start up with the launch of natural evolution based algorithms like genetic algorithm, ant bee search algorithm etc. and merged toward more rigorous approach by combining one algorithm with other such as artificial neural networks or ANN's and fuzzy logic as a black box technique to model systems and gained substantial interest in different areas of engineering. These are also known as hybrid techniques which consist of a frame- work of dynamic mass and energy balances, supplemented with fuzzy models. The hybrid models have shown that the use of fuzzy logic in hybrid modeling introduces flexibility, which enables the description of complex behavior with a pre defined, interpretable overall model structure. Araromi et al. [39] designed a continuous RD using hybrid Fuzzy Hammerstein (FH) model consisting nonlinear fuzzy model and linear state space model was then developed. The developed model was compared with linear autoregressive input exogenous (ARX) and nonlinear autoregressive input exogenous (NARX). Sumana et al. [40] investigated the use of gramian covariance matrices for sensor configuration in continuous multi component reactive distillation by applying extended Kalman Filter to obtain the instantaneous composition information from temperature sensors data of reactive distillation column. The sensors configurations were further evaluated by IAE criteria incorporating the measurements suggested by the state estimator. **Figure 3** shows general soft sensor control.

Soft computing or inferential computing is the most advanced stage of control schemes. At present, neural network is one of the most demanded intelligent controllers which works on the imitation of working of neurons in human brain. Wang et al. [41] have taken a case-based modeling program with an industrial example of distillation column. The basic features of this case base modeling described in brief were discreteness, nonlinearity, contradiction and complexity. Theyreported that neural network is promising in process control and fuzzy distributed neural network can be used to design a soft sensor for a high purity distillation column. Multilayer neural network was utilized when creating a system inverse neural model. In view of this, Zilkova et al. [42] have developed three-layer feed forward neural network with one hidden layer elected to approximate nonlinear function. The first subsystem served for desired current component reconstruction and second system

**Figure 3.**

*General structure of soft sensor control.*

serves for corresponding voltage components reconstruction for PWM converter. This was carried out using Simulink model. Simulation result verified the effectiveness of proposed controller. Akesson et al. [34] have taken control objective to keep the output close to a specified reference trajectory in such a way that large control signal variations were avoided and possible hard constraints on the state and inputs were satisfied. Main control objective of this work was to minimize the cost. For a data set, without having to compute the optimal MPC control signals by off-line optimizations. Raghavan et al. [43] have developed a Recurrent Neural Network (RNN) based inferential state estimation scheme for an ideal reactive distillation column. The performance of the estimator for both open loop and close loop was compared with that of kalman filter in terms of qualitative and quantitative indices and concluded that RNN has better level of inferential control over the conventional suggested methods. Prakash et al. [44] proposed an artificial neural network based nonlinear control algorithm for simulated batch reactive distillation column by considering a homogenously catalyzed esterification reaction which was controlled using ANN based state predictor. The open loop dynamic was presented in detail for this system and the proposed law was tested against gain scheduling controller to compare the performance. Bahar et al. [45] presented an Elman neural network to control the product composition from distillation column using temperature measurements inferentially. The main limitation of the neural network controller was that substantial offline computations may be needed in order to train it properly, and for some choices of cost functions it may not even be feasible to achieve satisfactory accuracies.

Other soft sensing techniques like genetic algorithm, particle swarm optimization, ant colony optimization etc. are although very old but are still being employed for inferential control which is based on natural evolution theory. Nithya et al. [46] have used real world experimentation in which pneumatic control valve is used to control the flow of water in and out of the tank. Using the black box modeling, the transfer function has been derived which was used to design a PI controller to find the values of gain and transfer function. Fuzzy logic controller has been designed for spherical tank considering nonlinear system. For tuning of the PI and Fuzzy Logic controller, genetic algorithm was used and the responses for servo and load disturbance were observed. Idris et al. [47] have considered the case of methyl acetate production in a continuous reactive distillation column in which tracking index term has been coined which was define as squared sum of differences between the predicted outputs and set points change over the prediction horizon. The control algorithms were applied in gPROMS against various tuning parameters and concluded that optimizing controller can be easily applied to a simulated complex model of reactive distillation to enable real time dynamic optimization.

#### *A Review on AI Control of Reactive Distillation for Various Applications DOI: http://dx.doi.org/10.5772/intechopen.94023*

Sujatha et al. [48] have considered a MIMO system defining integration of manipulated variable and control variable. Various novel techniques such as relative gain array, Niederlinski index, singular value decomposition, Morari Resiliency Index, dynamic relative gain array, hankel interaction index array, participation matrix and H2-norm were studied for the interactions and subsequently for input output pairing.

Differential evolution is one of the latest coined soft controlling techniques which aims at approaching toward global minimum. Subudhi et al. [49] have developed a differential approach based on a jumping rate fittest approach in which individual were selected from the union of current population and opposite population. The aim was to approach toward global minimum and then LM was used to move forward achieving fast convergence. LM is a gradient based algorithm to increase convergence speed. Opposition based learning improves the chance of starting with better initial population by checking the opposite solutions. Lu et al. [50] have developed a differential algorithm based stochastic search technique which is a powerful and global optimizer. The author has proposed a Modified Differential Evolution Fuzzy Neural Network (MDEFNN) which consists of a FNN identifier, a MDE estimator, a computation controller, and a hitting controller. Lawrynczuk et al. [51] have reported model predictive control strategy for a high purity, high pressure ethylene ethane distillation column. In this study, multi-layer perceptron neural network was applied with one hidden layer and a linear output. In the MPC-NPL algorithm, the nonlinear neural model was linearizing using eight repetitive steps such as estimation, approximation, control increment, and iteration. Kandanapitiya [52] has recently reported Modeling of Reactive Distillation for Acetic Acid Esterification. The mathematical model considered material balance equations, equilibrium relationships, summation equations and energy balance equations. The model was simulated for acetic acid and ethanol esterification reaction. Wierschem [53] has worked on design of Continuous Enzymatic Reactive Distillation with Immobilized Enzyme Beads. Based on kinetic and thermodynamic data, a detailed rate-based model of the ERD is developed. Simulation results and experimental data of the ERD setup are in good agreement. Zhang [54] has designed centralized and decentralized stochastic adaptive fuzzy output feedback control by using dynamic surface control method. Fuzzy systems are used to approximate unknown nonlinear continuous function. For the online process, the young's inequality norm of fuzzy basis vector is adjusted. Peng [55] has shown that two riccati equations were employed with relaxing the agent dynamics for uncertain nonlinear systems. The author has developed a cooperative output feedback adaptive control (COFAC). The NN identification of the individual uncertain dynamics is decoupled from the network topology, which is useful for practical implementations since the uncertain nonlinear dynamics can be suppressed by local NN. Cui [56] deals with the problem of adaptive decentralized NN control by combining Lyapunov-Razamikhin functional approach, minimum learning parameter algorithm and back stepping design technique. Here only adaptive parameter needs to be estimated for each subsystem which shows that all signal in close loop system are uniformly ultimately bounded in probability.
