ONLINE ESTIMATOR FOR DISTILLATION COLUMN USING ANN. Vijander Singh* Indra Gupta Puneet Gulati H.O Gupta

Transcription

1 ONLINE ESTIMATOR FOR DISTILLATION COLUMN USING ANN Vijander Singh* Indra Gupta Puneet Gulati H.O Gupta Department of Electrical Engineering Indian Institute of Technology Roorkee, Roorkee, Uttaranchal, India Abstract: This paper presents an ethernet-based data acquisition system, to provide a standard online estimator for distillation column. The data acquisition system is used to acquire various distillation parameters from any client PC connected to the network to estimate the distillate composition online. Secondary measurement technique is applied for inferential measurement of distillate quality. ANN based estimator using back propagation neural network technique is used for secondary measurements. The online estimated results are tested for single and double hidden layers with and ler with simulation results. Again for testing the robustness of the estimator, the estimated compositions are compared with experimental results using refractometer. The online estimator developed can be used by inferential control scheme for controlling distillate quality of distillation process. Copyright 2002 USTARTH Keywords: Distillation, Inferential control, Estimator, ANN and Data Acquisition.. INTRODUCTION Process automation and control has become an indispensable task in almost all areas of sciences and modern day industries. Industries need to collect data for control analysis and proposing new designs. In present work ethernet based data acquisition system for distillation column is developed. Ethernet provides a fast, reliable, low cost network. As the de facto networking standard, ethernet is well developed for a variety of networking and performance needs. To achieve the desired distillate quality, various transducers are used for instrumentation and control of distillation column. Using these transducers the data can be acquired and distillate composition can be estimated, which is the objective of distillation process. The importance of the study of distillation column dynamics was identified long back. Though the mathematical models required for simulating distillation process have been known since late 940 s. The equations derived for simulating distillation process are complex and therefore several simplifying assumptions has been made for building these models in order to obtain solution. The use of digital computers was introduced in 958 (Amundson and Pontinen in (958) to solve the distillation problem. * Corresponding author: vijaydee@gmail.com, Mob: , Fax: (Y.S.Choe and William L. Luyben in 987) worked on a rigorous dynamic model of distillation column. Most of the dynamic models assume the following two simplifications: ) Negligible vapor holdup and 2) Constant pressure. In 990 (Maurizio Rovaglio et al) solved the distillation column problem with the help of rigorous model. The control of many industrial processes is difficult because online measurement of product quality is complicated. Sometimes the instrumentation is either very expensive and/ or measurement lags and sampling delays make impossible to design an effective feedback control system. (R. Weber and C. B. Brosilow in 972) cited one solution to this problem by using secondary measurements in conjunction with a mathematical modeling of the process to estimate product quality. The method includes procedures for selecting the available output measurement to get an estimator, which is relatively insensitive to modeling error and measurement noise. The estimator developed for control of multicomponent distillation column is based on temperature, reflux and steam flow measurements. (B. Joseph and Coleman Brosilow in 978) presented a method for designing an estimator to infer immeasurable product qualities from secondary measurements. The secondary measurements are selected so as to minimize the number of such measurements required to obtain an accurate estimate. The application of design procedures to design a static inferential control system to control product composition is described. (E.Q. Marmol et al in 99) applied an observer for

2 r control of multi-component batch distillation. The observer is a dynamic model that estimates the state variables of a process. Some state variables that are not accessible for measurement or measured only as a combination of some state variables, for example temperature of a tray is the result of mole fraction of all the components on that particular tray and pressure. It was shown that state vector of a linear system can be reconstructed from the observation of systems input and output. In the distillation column application, the state variables are compositions in every stage. (In 992 E.Q. Marmol and William L. Luyben) presented an inferential model based control of multi-component batch distillation. ANNs are commonly used as inferential model in distillation columns. (Phiroz Bhagat in 990) discussed briefly the neural networks. Two examples were taken to demonstrate their practical application, these involved CSTR s. (In 994 A. J. Morris et al) examined the contribution of various network methodologies to the process modeling and control toolbox. (Vijander Singh et. al. in 2005) developed the ANN based offline estimator for distillation column. COMPRESSED AIR P AIR COMPRESSOR Temperature signal from estimator PID Controller COOLANT WATER PRESSURE REGULATOR HEATER FEED TANK Vent P FT REBOILER r PT ACV MCV STEAM TRAP PD CONDENSER LT PT R BOOM PRODUCT TANK r REFLUX DIVIDER ANN Based Estimator (T B, T,...T N D ) MCV LT TOP PRODUCT TANK xd MCV Temperature signal to PID controller INDEX Temperature transmitter FT Flow Transmitter PT Pressure Transmitter LT Level Transmitter P Pump PD Power Driver ACV Automatic control valve MCV Manual control valve r Rotameter Figure Schematic Diagram of Distillation Column In the present work, a complete online data acquisition scheme based on ethernet is developed for laboratory set-up of distillation column and an ANN based estimator for distillation column is designed to estimate the distillate composition online. The developed estimator is tested with the single and double hidden layers with the simulation results for an experimental setup of distillation column. In the experimental set-up of the distillation column as shown in Figure, temperature transducers on each tray, inlet and outlet of condenser and reflux drum, pressure transducer at the bottom and the top of distillation column, level transducer in the reboiler and flow transducer at the feed path are employed. Rotameters are provided for controlling the feed flow, bottom product and cooling water. 2. DATA ACQUISITION VIA ETHERNET The Internet continues to become more integrated into our daily lives. This is particularly true for scientists and engineers, because designers of development systems view the Internet as a costeffective worldwide standard for distributing data. Data is acquired from the distillation column with the help of various transducers attached to it. The outputs of these transducers are connected to the National Instruments Field Point Modules. The data is transferred to the client terminal via ethernet with the help of these modules. So the data acquired from the field point modules can be viewed from any terminal by just connecting the client terminal to the server of the field point modules. These are network modules consisting of a server module and several I/O modules to send and receive data from/to the process side and the client side. These server module and I/O modules form a bank can be accessed by any number of client computers. These modules can automatically detect the speed of the connection and configures itself accordingly. Field Point software (provided by National Instruments) is used to create the configuration (.iak) file. It communicates with field point hardware and, read and write data to/from the modules. The network modules provided by National Instruments can be used effectively to acquire data from the remote site. The connection is established between the network module server and client PC by creating the data sockets and the OPC server present in the client system, which transfers the data from the process site to the client PC. 3. SECONDARY MEASUREMENT TECHNIQUES Inferential Measurement is a powerful and popular methodology that allows product quality to be inferred from other easily measurable plant variables. This is due to the reason that the direct measurement of product quality is difficult. The behavior of any process is indicated by the states of 2

3 output variables, which are dependent on the operating conditions and the adjustments made to the process. However, productivity is quantified by a subset of these output variables; normally the specifications upon which the product is sold, e.g. purity, physical or chemical properties. These primary variables are often difficult to measure online. Inferential measurement systems are designed to overcome such measurement problems. The other outputs, (for example temperatures, flows and pressures) are called secondary variables and these are easily measured online. Due to the nature of chemical and process engineering systems, the states of secondary variables reflect the states of primary variables. Thus it is possible to use the readily available secondary variables to infer the quality of state or primary variable. In developed inferential measurement systems, the objective is to model the relationship between a primary output and secondary measurement variables. Then the model can be used to estimates the distillate quality which is difficult to measure directly. If the distillate quality is sufficiently accurate, then inferred states of primary outputs can be used as feedback for automatic control and optimization of distillation process. The primary objective in distillation column is to get the distillate product of the desired composition. For this purpose it is very important to measure the top product composition. But measuring the composition each time using refractometer is a very tedious and time consuming. Also using online sensors for distillate quality implies very high costs and time delays. So, considering above points, inferential measurement techniques have been used and estimator has been designed to measure the composition online without any delay. For this purpose the ANN based estimator is developed and coupled with data acquisition software for online estimation. The temperature profiles of all the trays have been used as the secondary variables and are used as inputs for inferring distillate composition. 4. PID CONTROLLER USED IN DISTILLATION COLUMN PID Controller used in laboratory set up of distillation column as shown in Figure, is a digital indicating PID Controller model no. UT320 manufactured by Yokogawa, for keeping the temperature of the particular connected tray constant throughout so that distillate composition can be controlled. However in the present work it is used for two purposes: For supplying variable heat inputs to the reboiler for collecting patterns for ANN training by operating it in manual mode. In the reboiler of distillation column one heater of 4 kw and two heaters of 2 kw each exist and can be used according to requirement. Out of these three heaters PID controller is connected to 4 kw heaters to control the heat input to reboiler and rest 2 kw heaters are used directly without PID control. Base heat input of 4kW is always supplied and the rest 4 kw is varied using PID controller. Therefore heat is varied between 4-8 kw in small steps of 0.25 kw. The interfacing of PID controller is shown in figure 2. For keeping the temperature of the particular tray constant so that constant temperature patterns are also included by operating it in automatic mode. So the PID controller maintains the tray temperature at the desired value. The PID controller is interfaced with the computer using a RS-485/232 converter and is connected with the serial port of the PC. Set Point from User Tray Temperature PID Controller Controller Signal SCR Based Temperature Controller Reboiler heater Figure 2 PID controller in distillation column 5. GENERATION OF PAERNS FOR ANN Experiments are performed by varying the heat input to the reboiler by operating the PID controller in the manual mode With the help of ethernet based data acquisition software developed all the tray temperatures, top and bottom pressure and also percentage of flow were stored in the database. Also distillate composition was noted for each observation. Also some readings are taken by keeping the temperature of Tray constant at 85 C and 90 C with the help of PID controller by operating the PID controller in auto mode. So total 450 patterns were collected and used for training the network and some readings are kept for testing the network. The purpose of taking four combinations is to choose the optimum architecture out of these to be actually employed for online estimation. For obtaining optimum results large number of epochs have been performed using the back propagation algorithm. 3

4 6. ARTIFICIAL NEURAL NETWORKS AND ONLINE ESTIMATOR Artificial neural networks are used for inferential measurement and thus building the online estimator. As it depends only on example based learning, requires no explicit relationship between the input and output and can efficiently model non-linear behavior. Two models of Neural Networks have been implemented with one and two hidden layers respectively. Also two types of patterns have been used for training the neural network, one and the other with PID control. So in all four combinations of ANN, the online estimator is implemented. Figure 3 shows the architecture with hidden layer. Similar architecture is used for 2 hidden layers also. The number of hidden neurons is chosen by hit and trial, and so are the values of momentum and learning rate. Number of hidden layers are restricted to two as it is clear that two layers are sufficient to model many complex relationship and if hidden layers are increased then the neural network adopts the method of memorizing instead of learning. After performing the experimentation it is observed that ANN architecture with single hidden layer is sufficient for inferential measurement for the distillation process of binary mixture of methanol and water and also gives satisfactory results when trained with data obtained ler. PID controller does not add any additional advantage for the training. 2.0 lit/hr. The feed rate was at 2.5 lit/hr to keep the quantity of mixture as low as possible and the reflux ratio is kept at 4.0 to obtain the distillate quality in the purest form. After performing rigorous training, the estimator developed is tested with some test patterns, which are not included during the training process. There are 33 test patterns ler and 47 test patterns with PID controller. Figures 4(a-d) show the result of these testing a. Distillate composition with layer and b. Distillate composition with 2 Layer and. Fig.3 Artificial Neural Network Model 7. EXPERIMENTATION AND RESULTS At the time of starting the laboratory setup of distillation column several valves have to be fixed at some set value. The value of rotameters for cooling water is set at 00lit/hr to maintain the temperature of the distillate (methanol) below its boiling point. The bottom product rate is kept at c. Distillate composition with layer and PID control 4

5 > Readings ---> 4d. Distillate composition with 2 hidden layer and with PID control 5b. Distillate composition with 2 layers and Figure 4 Testing Results for Estimator It is observed from figures 4(a-d) and 5(a-d) for single and double hidden layers with and ler that the estimated results are in good agreement and harmony with the simulation and experimental results for the developed estimator respectively. It is also observed that using double hidden layer with PID controller the estimated composition of distillate does not improve significantly. This is due to the reason that the estimator estimates the composition of binary mixture, which does not have complex nonlinearities. -> Readings -----> 5a. Distillate composition with layer and ----> Readings ----> 5c. Distillate composition with layer and with PID control -----> Readings > 5 d. Distillate composition with 2 layer and with PID control Figure 5 Results of Online Estimation 8. CONCLUSIONS In the present paper, a complete online data acquisition scheme based on ethernet is developed and an online estimator for distillation column is developed. It has been effectively utilized to acquire data both for collecting the patterns and for online estimation of distillate composition. Rigorous experimentation has been carried out to collect patterns for data based 5

6 modeling adopted in the present case. ANN has been chosen as the architecture for building the estimator for inferential measurement. The developed estimator is trained successfully and then tested for different combinations hidden layers and PID controller. The test results show the closeness of the estimated values with the simulation and experimental values of distillate composition with and ler. These are then coupled with data acquisition software to estimate the distillate composition online. It observed from figures 4 and 5 that the developed estimator for the distillation column with single layer of hidden neurons and without PID controller gives satisfactory results. Further it is observed that increasing the number of hidden layers and including PID controller does not add any improvement in the performance of the developed estimator for binary mixture. This is due to the reason that the developed estimator estimates the distillate composition of a binary mixture, which does not have complex nonlinearities. REFERENCES A.J. Morris, GA. Montague and M.J. willis (Jan 994), Artificial Neural networks: Studies in process modelling and control, Trans. I. Chem Engg, Vol 72, Part A, pp Babu Joseph and Coleman B. Brosilow, (978) Inferential control of process: part-i Steady State Analysis and Design, Part-2 The structure and Dynamics of Inferential Control Systems, Part-3 Construction of Suboptimal Dynamic Estimators, AIChE J, Vol. 24, No 3, pp Enrigue Quintero-Marmol, William L. Luyben and Christos Geogarkis (99), Application of an extended Luenberger Observer to the control of Multi-Component Batch Distillation, Ind. Engg. Chem. Res,. Vol 30, No.8, pp Enrigue Quintero- Marmol and William L. Luyben (992), Inferential model based control of multi-component Batch distillation, Chemical Engineering Science, Vol 47, pp Mourizio Rovaglio, Eliseo Ranzi, Giuseppe Biardi, Marco Fontana and Rosa Domenichini (April 990), Rigorous Dynamic and feed forward control design for distillation process, AIChE Journal, Vol 36, No.4, pp Neal R. Amundson and Arlene J. Pontinen (May 958), multicomponent distillation calculations on a large digital computer, Industrial and Engg. Chemistry, Vol 50, No. 5, pp Phiroz Bhagat (August 990), An Introduction to Neural Nets, Chemical Engg. Progress, pp Richard Weber and Coleman B. Brosilow (May 972), The use of secondary measurements to improve control, AIChE Journal, Vol 8, No.3, pp Vijander Singh, Indra Gupta, and H.O Gupta (July 2005), ANN Based Estimator for Distillation Column-Inferential Control, Chemical Engineering and Processing, Vol 44, No. 7, pp Young-Soon Choe and William L. Luyben (987), Rigorous Dynamic models of distillation columns, Industrial Engg. Chemistry Res. Vol 26, No. 0, pp