Neural-based Monitoring of a Debutanizer Distillation Column L. Fortuna*, S. Licitra, M. Sinatra, M. G. Xibiliaº ERG Petroli ISAB Refinery, 96100 Siracusa, Italy e-mail: slicitra@ergpetroli.it *University of Catania, DEES, Viale A. Doria 6, tel 39957382307, 95100 Catania, Italy ºUniversity of Messina, Dept. of Mathematics Contrada Papardo, Salita Sperone 31, 98166 Messina, Italy e-mail: mxibilia@ingegneria.unime.it Keywords: Modelling, Neural networks, Virtual sensors, Distillation columns Abstract In this paper a neural approach to distillation columns modelling is described. In particular a Debutanizer colums is considered and a real-time estimate of the butane percentage (C4) in the bottom draw (C5) is obtained by a NARMAX model implemented with a Multi-Layer Perceptron. The analyser of the C4 in C5 percentage used at present, provides a measure after a great and unknown delay, and is therefore not suitable for closed loop control purposes. A neural-based model, acting as a virtual sensor, can therefore represent a suitable strategy in getting a real-time estimation of the C4 in C5 concentration. Neural networks are used both to evaluate the delay of the analyser and to provide the desired real-time estimate of the C4 flow in the bottom draw of the debutanizer, overcoming the analyser s delay. To obtain more accurate results the model is built so that the measured output is used as an input of the model together with the predicted one, suitably delayed. The neural NARMAX model has been determined by using an appropriate set of measurements performed on a plant operating in Sicily (Italy) and is now working on the plant. A comparison between the estimated output and the analyser's measures confirms the validity of the proposed approach.
1. Introduction. In the last years an ever growing interest has been given to the production quality standards and to the pollution phenomena in industrial environment. Particular attention has been devoted to petrol-chemical area due to the high risks deriving from the production process and to the high level of quality standard imposed by government laws. In this context, due to the complexity of the considered systems, in general highly non-linear, and to the high number of variables and disturbances, the analytical modelling strategies cannot be easily applied. Great improvements can be obtained approaching the involved measurements and control problems with non conventional techniques, as for example Soft Computing, which allow to obtain non-linear dynamical models joining neural networks, fuzzy logic and genetic algorithm, exploiting human knowledge and measurements performed on the system. In this paper the first part of this monitoring and control strategy is described. In particular a virtual sensor, implemented via a MLP neural network is described. The system obtained can be used as virtual sensor, as parts of control loop or to validate the sensors output, detecting the occurrence of faults in the measurement system. Such an approach allows a great economical advantage, in that a lower number of measuring devices is needed, the measures can be easily validated and a more tight control on the production quality can be obtained. A Soft Computing approach in modelling of distillation columns has been used in [1] and [2] where the benzene percentage of a Splitter Benzene column has been considered. In particular 3 different types of model have been determined: a one-step predictive model to obtain a real-time estimate, in spite of the 15 minutes delay introduced by the chromatograph, a non linear MA model, able to give an approximative value of the benzene percentage in case of faults of the chromatograph, and a three-step predictive NARMAX model, used to share the multistream chromatograph among 3 different columns. The application considered in this paper is the butane concentration neural modelling in the bottom stream of a Debutanizer distillation column operating in the ERGPETROLI ISAB Refinery (Siracusa, Italy). The butane concentration is at present indirectly measured by an analyser which is connected to a different column, the measure is therefore given with a great and unknown delay and cannot be efficiently used in a control loop.
As it will be explained in the following, the proposed neural model allows to obtain a real-time estimate of the percentage of C4 in C5, thus overcoming the delay introduced by the present measurement system. 2. The distillation process: debutanizer column. The distillation process allows to separate the mixture coming from the Deethanizer in several components through a sequence of vaporisation and condensation steps [4]. It can be realized in a distillation column, as represented in Fig. 1. TI 040 PRC 011 E107 A/B FRC 015 D104 LRC 008 FRC 018 G.P.L. Splitter 34 22 Deethanizer P102 A/B P103 A/B 9 TRC 004 TI 037 FRC 016 T102 TI 036 E108 A E108 B FRC 021 Nafta slpitter Fig. 1 : The Debutanizer The column has two draws: top and bottom. The overheads from the column are condensed against cooling water (E107A/B) and enter the debutanizer reflux drum (D104). Liquid from the reflux drum is either returned to the column as reflux (F015RC) or pumped to the C3/C4 splitter (F018RC). The flow to the C3/C4 splitter is normally cascaded from the drum level (L008RC). C3/C4 product flow is typically 50-70 m 3 /Hr versus a reflux flow of around 200m 3 /Hr. The column pressure (P011RC) usually manipulates the
flow through the water condensers (E107A/B), but, in case of very high pressure, the controller can also open the blow down valve (F037R). Temperature at the top of the column is measured (T040I); temperature is measured also in the middle of the column (T004RC) and its variation controls the flow of the vacuum residue against witch the bottom of the column is reboiled (E108A/B) and hence influence the two temperatures (T036I and T037I) measured at the bottom of the column. Liquid is withdrawn from the column under flow control (02F021RC) and routed to the Naphtha splitter. The first step in realizing a virtual sensor of the C4 flow in the bottom draw of the column is to select the correct set of input variables. This has been done by using the knowledge of the experts operating on the system. After an accurate analysis of the involved variables, based on the process knowledge, the most important have been selected as model inputs as follows: u1: (T040I) Top temperature; u2: (P011RC) Top pressure; u3: (F015RC) Top reflux; u4: (F018RC) Top drawn flow; u5: (T004RC) Middle temperature; u6: (T036RC) Bottom temperature; u7: (T037RC) Bottom temperature. The model output is the flow of C4 in the bottom draw: y: (F_C4). Historical plant data covering an operating period of two months were collected from refinery database to build the neural model. The trends of these variables are shown in Fig. 2 The C4 composition in the Debutanizer bottom is not measured at the bottom draw, because the C4 analyser is installed on the overheads of the Naphtha Splitter. We can suppose that the analyser measures all the flow of C4 that comes out from the bottom of the Debutanizer because in the Naphtha Splitter C4 is the lighter
component so it will go unbroken in the top of the Naphtha Splitter. The measure provided from the analyser has therefore a great delay with respect to the actual output. Fig.2 The seven input variables and the output of system This delay should be properly estimated in order to use the analyzer output data in building the neural model. From our knowledge about the plant's structure, we can only suppose that the delay of the analyser is in the range [45 min - 1h:30']. 3. Neural Modelling and Numerical Results As well known, a lumped non-linear dynamical system can be represented by a NARMAX model of the following form [3], [6], [7]: y(t) = f [ y1( t-1 ), y1( t-2 ),, y1(t-n 1 ),, y2(t-1),,y2(t-n 2 ),yn(t-1),, yn(t-n n ), u1( t-1), u1(t-m 1 ),, um(t-m m )] (1) where t is the discrete time, y=[y1,,yn] is the n-dimensional output vector, n i (i=1,,n) is the number of delays of the i-th output variables, u=[u1,,um] is the m-dimensional input vector and m i (i=1,,m) is the number of delays of the i-th input. A model is therefore determined by the non linear function f(.) and by the number of input and output delayed samples.
Due to the approximation capabilities of MLP neural networks [5] the function f(.) can be suitably represented by a MLP if a sufficient number of measured I/O samples is used to train the network and the correct number of I/O delays are considered [6]. In the considered application the correct delay between the actual and measured output (TD) should be also determined, in order to arrange correctly the measurements and to obtain a suitable training data set. The number of delayed samples to be used for each I/O variable has been determined on the base of both the knowledge of the plant operators and a trial and error phase. A large set of neural networks, differing for the value of TD, the input and output variables delays and for the number of hidden neurons has been trained. Best results have been obtained by considering TD=45 min. As an example of the comparison performed, in Fig 3 the difference between the actual output and the model's output is shown when two different values of TD (45 min and 60 min) are considered to arrange the data set. The model inputs and the network structure are identical in both cases. Fig 3 Difference between actual output and model output when 2 different delays are considered to arrange the data set. Over about 1000 measured data (sampling time is 15 min.), 800 have been used to train the networks while the remaining 200 have been used to test their performance. It can be seen that the performance of the
two networks, one trained supposing TD=45 min and the other with TD=60 min, are very different as regards the 200 checking data, indicating that 45 min is a correct estimate of TD. At the end of this trial and error phase good results are obtained using the following input variables: FC_4 * (k+1) = f [T040(k), P011(k), F015(k), F018(k), T004(k-3), T004(k-2), T004(k-1), T004(k), T036(k), T037(k), FC_4 * (k), FC_4 * (k-1)] where k is the discrete time, FC_4 * is the estimated value of the output and the other input values are given by the corresponding measurement instruments. In order to further improve the model performance the following remark can be considered: since the analyser has a sampling time of 15 minutes, a model with only two delayed output samples as inputs, used on-line, cannot use the analyser measured values, because the analyser has not yet provided the two needed samples (TD is 45 min, corresponding to 3 prediction step). So the neural network has to iteratively reconstruct the two estimated output value used in the input vector and the analyser measure is not used at all. Therefore a new model with four delayed output samples has been considered: two samples came from the analyser, while the other two samples (the most recent ones) are computed by the network itself. A new set of 14-x-1 neural network has been therefore trained, with a larger set of data in order to determine the correct number of hidden neurons. About 1500 data have been used to train the network and a different set of 500 data has been used to test the results. The number of hidden neurons has been chosen with a growing strategy, checking the possible occurence of overlearning by computing the output MSE both with the learning and testing data sets. Fig 4 shows the neural network output and the analyser output; the difference between these two trends is shown too. The obtained virtual sensor is presently implemented on the refinery control and monitoring system. Fig. 5 represents the actual versus the model output obtained after the implementation of the model. As can be observed the obtained performance are satisfactory. However a periodic tuning of the network weights has to be done to take into account that the plant is sometimes stopped and restarted from a different working point.
Fig 4 Dynamic model output and real output and their difference Fig. 5 On-line performance of the virtual sensor
4. Conclusions In the paper the advantages introduced by the neural network modelling approach to a petrol chemical plant are outlined. In particular bottom draw butane concentration of a distillation column is predicted in real time as a function of a suitable set of input variables via a MLP. A delayed value of the output variable, measured by an analyser located on a different column is used to improve the model performance. The choice of the input variables has revealed of fundamental importance for the efficiency of the model and has been made based both on process knowledge and on trial-and-error procedures. The dynamical model allows a real-time control of the bottom product of the debutanizer. The results obtained are satisfactory and suggest the application of neural modelling in solving a number of measurement and control problems in petrolchemical processes. A further improvement can be introduced by using also hybrid metodologies, based on neuro-fuzzy networks and genetic algorithm both in realising virtual sensors and sophisticated control algorithm. References [1] C. Bozzanca, S. Licitra, L. Fortuna, M.G. Xibilia, Neural networks for benzene percentage monitoring in distillation columns, Proc. of Soft Computing (SOCO) 99, pp. 391-395, Genova, Italy, June 1-4 1999. [2] M. Bucolo, L. Fortuna, M. Sinatra, S. Graziani, Neuro-fuzzy modelin in petrol-chemical industry, Proc. of the 7-th Mediterranean Conf. On Control and Automation, JUNE 28-30, Haifa, Israel, 1999. [3] L. Ljung, T. Soderstrom, Theory and Practice of Recursive Identification, Cambridge, Mass: MIT Press, 1983 [4] R. Perry, Don Green, Perry s chemical engineers handbook, Mc-Graw Hill, 1984 [5] G. Cybenko, Approximation By Superposition of a Sigmoidal Function, Mathematics of Control, Signals, and Systems, Springer-Verlag, pp. 303-314 [6] S. Chen, S.A.Billings and P.M.Grant, Non-Linear system identification using neural networks, Int. Journal of Control, vol. 51, No. 6, 1191-1214,1990. [7] K. S. Narendra, K. Pathasarathy, Identification and Control of Dynamical System Using Neural Networks, IEEE Trans. on Neural Networks, Vol. 1, No. 1, 1990.