Process modeling and optimization of mono ethylene glycol quality in commercial plant integrating artificial neural network and differential evolution

Transcription

1 From the SelectedWorks of adeem Khalfe Winter December 7, 2008 Process modeling and optimization of mono ethylene glycol quality in commercial plant integrating artificial neural network and differential evolution adeem Muhammed Khalfe Available at:

2 CHEMCO, Process modeling and optimization of mono ethylene glycol quality in commercial plant integrating artificial neural network and differential evolution S.K. Lahiri Department of Chemical Engineering, ational Institute of Technology, Durgapur, India adeem M. Khalfe Jubail United Petrochemical Company, SABIC. Sunil Kumar Sawke Jubail United Petrochemical Company, SABIC. ABSTRACT This paper presents an artificial intelligence based process modeling and optimization strategies, namely artificial neural network differential evolution (A-DE) for modeling and optimization of ultraviolet (UV) transmittance of mono ethylene glycol (MEG) product. UV transmittance is one of the most important quality variable of MEG that has impact on the polyester product quality. UV transmittance measures the presence of undesirable compounds in MEG that absorb light in the ultraviolet region of the spectrum and indirectly measures the purity of MEG product. They are in trace quantities in the ppb ranges and primarily unknown in chemical structure. Thus, they cannot be measured directly. Off-line laboratory method for MEG UV measurement is common practice among the manufacturer, where a sample is withdrawn several times a day from the product stream and analyzed by time consuming laboratory analysis. In the event of a process malfunction or operating under suboptimal condition, the plant continues to produce off-spec product until lab results become available. It results in enormous financial losses for a large scale commercial plant. In the present paper a soft sensor was developed to predict the UV transmittance on real time basis and an online hybrid A-DE technique was used to optimize the process parameters so that UV is maximized. This paper describes a systematic approach to the development of inferential measurements of UV transmittance using A regression analysis. After predicting the UV accurately, model inputs are optimized using DEs to maximize the UV. The optimized solutions when verified in actual commercial plant resulted in a significant improvement in the MEG quality. Keywords: A, DE, modeling & optimization

3 . ITRODUCTIO: Recently Monoethylene glycol (MEG) has emerged as most important petrochemical product as its demand and price rises considerably in last few years all over the world. It is extensively used as a main feed for polyester fibre and polyethylene tere-phthalate plastics production. UV is one of the most important quality parameter of MEG and its represents indirectly the impurities level such as aldehyde, nitrogeneous compound and iron in the MEG product. In laboratory, MEG product sample is exposed to UV light of different wavelengths (220, 250,275 and 350 nm) and how much percentage of the UV light is transmitted through the MEG sample is measured. UV transmittance measures the presence of compounds in MEG that absorb light in the ultraviolet region of the spectrum. These undesirable compounds are in trace quantities in the ppb ranges and primarily unknown in chemical structure. Samples showing higher transmittance are considered to be of a greater quality grade. In Glycol plant the MEG is drawn off from MEG Column as product, its UV transmittance is affected by many things such as impurity formation in upstream ethylene oxide reactor, impurity formation and accumulation in MEG column bottoms due to thermal degradation of glycol, non removal and accumulation of aldehyde in the system etc. Because these UV deteriorating impurities are in ppb ranges, they are very difficult to detect during MEG production process and they have hardly any effect on process parameters.that s why it is very difficult for any phenomenological model for UV prediction to succeed in industrial scenario. ormally online UV analyzers are not available to monitor product MEG UV analysis in ethylene glycol plant, so offline methods for MEG quality control is common practice among the manufacturer, where a sample is withdrawn from the process and product stream for laboratory analysis several times a day and analyzed by time consuming laboratory analysis. In the event of a process malfunction or operating under suboptimal condition, the plant will continue to produce off-spec product until lab results become available. For a big world class capacity plant this represents a huge amount of offspec production results in enormous financial losses. This necessitates the online UV sensors or analyzers which can gives UV continuously on real time basis. Accurate, reliable and robust UV soft sensors can be a viable alternative in this scenario. Making of UV soft sensor is not an easy task as rigorous mathematical model for MEG product UV is still not available in literature which can predict UV transmittance to minimize the dependency on lab analysis. The comprehensive process model is expected to take into account the various subjects, such as chemistry, chemical reaction, UV deteriorating compound generation and accumulation which consequently become very complex. Industry needs this mathematical model to predict MEG UV on real time basis so that process parameters can be adjusted before the product goes off specification. To develop such model from basic principles of chemical engineering is very difficult due to unknown reactions taking place. In the last decade, As have emerged as attractive tool for nonlinear modeling especially in situations where the development of phenomenological or conventional regression models becomes impractical or cumbersome. The advantages of an A-based model are (i) it can be constructed solely from the historic process input-output data (ii) detailed knowledge of the process phenomenology is unnecessary (iii) a properly trained model possesses excellent generalization ability owing to which it can accurately predict outputs for a new input data set [2]. 2

4 Once an A based process model is developed, it can be used for predicting the MEG product UV s. The model can be utilized for UV soft sensor development and can be interfaced with online DCS and continuous monitoring can be achieved to yield the better process control. Once an A based process model is developed, it can be used for process optimization to obtain the optimal values of the process input variables that maximize the MEG product UV. In such situations, an efficient optimization formalism known as Differential Evolution which is lenient towards the form of the objective function can be used []. The DEs were originally developed as the genetic engineering models mimicking population evolution in natural systems. Specifically, DE like genetic algorithm (GA) enforce the survival-of the- fittest and genetic propagation of characteristics principles of biological evolution for searching the solution space of an optimization problem. The principal features possessed by the DEs are: (i) they require only scalar values and not the second- and/or first-order derivatives of the objective function, (ii) capability to handle nonlinear and noisy objective functions, (iii) they perform global search and thus are more likely to arrive at or near the global optimum. In the present paper, A formalism is integrated with DE to arrive at modeling and optimization strategies. The strategy (henceforth referred to as A-DE ) use an A as the nonlinear process modeling paradigm, and the DE for optimizing the input space of the A model such that an improved process performance is realized. To our knowledge, the hybrid involving A and DE is being used for the first time for chemical process modeling and optimization. In this study, the A-DE strategy have been used to model and optimize the MEG product UV for a commercial plant The optimized operating conditions leading to maximized UV of the product (MEG). The best sets of operating conditions obtained thereby when subjected to actual plant validation indeed resulted in significant enhancements in UVs. 2. Hybrid A and DE Based Modeling eural networks are computer algorithms inspired by the way information is processed in the nervous system. 2. etwork Architecture: The back propagation algorithm[2] assumes a feed forward neural network architecture (as shown in fig.) where nodes are partitioned into layers. The lower most layer is the input layer numbered and the topmost layer is the output layer numbered. Back propagation addresses networks containing one or more hidden layers. Hidden nodes do not directly receive inputs from nor send outputs to the external environment. Input layer nodes merely transmit input values to the hidden layer nodes and do not performs any computations. The number of input nodes equals the dimensionality of input patterns and the number of nodes in output layer is dictated by the problem under considerations. Each hidden node and output node applies the activation function to its net input. ormally there are three types of activation function reported in literature namely sigmoid function, tan hyperbolic function and linear function. 3

5 Fig. Architecture of feed forward network with one hidden layer 2.2 Training Training a network consists of an iterative process in which the network is given the desired inputs along with the correct outputs for those inputs. It then seeks to alter its weights to try and produce the correct output (within a reasonable error margin). If it succeeds, it has learned the training set and is ready to perform upon previously unseen data. If it fails to produce the correct output it rereads the input and again tries to produce the correct output by adjusting the weights. 2.3 Generalizability eural learning is considered successful only if the system can perform well on test data on which the system has not been trained. This capability of a network is called generalizability. Given a large network, it is possible that repeated training iterations successively improve performance of the network on training data e.g. by memorizing training samples, but the resulting network may perform poorly on test data (unseen data). This phenomenon is called over training. The proposed solution is to constantly monitor the performance of the network on the test data. In literature it is proposes that the weight should be adjusted only on the basis of the training set, but the error should be monitored on the test set. Here we apply the same strategy: training continues as long as the error on the test set continues to decrease and is terminated if the error on the test set increases. Training may thus be halted even if the network performance on the training set continues to improve. 2.4 DE Based Optimization of A Models Having developed an A-based process model, a DE algorithm is used to optimize the - dimensional input space (x) of the A model. Differential Evolution (DE), an improved version of GA, is an exceptionally simple evolution strategy that is significantly faster and robust at numerical optimization and is more likely to find a function s true global optimum. The optimization objective underlying the DE-based optimization of an A model is defined as: Find the -dimensional optimal decision variable vector, x* =[x *, x 2 *. x n *] T representing optimal process conditions such that it simultaneously maximizes process outputs, y. In the DE procedure, the search for an optimal solution (decision) vector, x*, begins from a randomly initialized population of probable (candidate) solutions. The solutions are then tested to measure their fitness in fulfilling the optimization objective. Implementation of this DE algorithm and looping generates a new 4

6 population of candidate solutions, which as compared to the previous population, usually fares better at fulfilling the optimization objective. The best vector that evolves after repeating the above described loop till convergence forms the solution to the optimization problem. (refer fig 2) Fig2: Flowchart for DE based optimization of A model 5

7 3. Case Study of Mono Ethylene Glycol Product UV Transmittance Fig 3 describes a brief process description where Glycol (90%) and water solution (0%) fed to the drying column to remove the water from drying column top. The bottom of drying column fed to MEG column to distil MEG from heavier glycols ( namely diethylene glycol and triethylenen glycol). MEG product (99.9 % wt purity) is withdrawn from the MEG column below the top packing bed. An overhead vapor purge of up to 0 % of the product is taken overhead to purge the light compounds. Fig 3 shows the location of input parameters from drying column and MEG column which were used to build the model of UV. Fig: 3 Process Flow diagram of Drying and MEG column 3. Development Of The A Based Correlation The development of the A-based correlation had been started with the collection of a large databank. The next step was to perform a neural regression, and to validate it statistically. 3.2 Collection of Data The quality and quantity of data is very crucial in A modeling as neural learning is primarily based on these data. Hourly average of actual plant operating data at steady state was collected for approximately one year. Data was checked and cleaned for obvious inaccuracy and retains those data when plant operation was in steady state and smooth. Finally 6273 records are qualified for neural regression. This wide range of database includes plant operation data at various capacities starting from 75% capacity to 0% of design capacity. 3.3 Identification of Input and Output Parameters The column performance was monitored in terms of output variable namely UV. Based on the operating experience in glycol plant, all physical parameters that influence UV are put in a so-called 6

8 wish-list. Out of the number of inputs in wish list several sets of inputs were made and tested via rigorous trial-and-error on the A. The above mentioned criteria were then used to identify the most pertinent set of input groups. Based on the above analysis, the nine input variables (in table ) have been finalized to predict UV. Table Input and output variable for A model Input Variables Reflux Ratio (Product flow / Reflux flow) Reflux Flow (MT/Hr) MEG Column Top Pressure (mmhg) MEG Column Condenser Pressure (Barg) MEG column control temperature (Deg C) MEG column feed flow (MT/Hr) Drying column control temperature (Deg C) Drying column bottom temperature (Deg C) Crude Glycol reprocessing flow. Output variables Mono ethylene glycol UV 3.4 eural Regression For modeling purposes, the reaction operating conditions data ( see table) can be viewed as an example input matrix (X) of size (6273 X 9), and the corresponding UV data as the example output matrix (Y) of size (6273 X ). For A training, each row of X represents a nine-dimensional input vector x = [x, x2 x9] T, and the corresponding row of matrix Y denotes the one-dimensional desired (target) output vector y = [y] T. As the magnitude of inputs and outputs greatly differ from each other, they are normalized in 0- scales.to avoid over training phenomena described earlier, 80% of total dataset was chosen randomly for training and rest 20% was selected for validation and testing. It has been reported that multilayer A models with only one hidden layer are universal approximators. Hence a three layer feed forward neural network (like Fig.) is chosen as a regression model. As there is no previous idea about the suitability of the particular activation function, all the three activation function (sigmoid, tan hyperbolic and linear) are chosen in all combinations for both hidden layer and output layer. The purpose is to find out which combination gives lowest error. The number of nodes in the hidden layer is up to the discretion of the network designer and generally depends on problem complexity. With too few nodes, the network may not be powerful enough for a given learning task. With a large number of nodes (and connections), computation is too expensive and time consuming. In the present study, the optimum number of nodes is calculated by trial and error method. The statistical analysis of network prediction is based on the following performance criteria:. The average absolute relative error (AARE) should be minimum 7

9 AARE = ypredicted y experimental ( ) y experimental R= i= 2. The standard deviation(σ)should be minimum σ = [ ( ypredicted y experimental y erimental AARE]^ ( i) ( i) ) / exp ( i) 2 3. The cross-correlation co-efficient (R) between input and output should be around unity. i= ( y experimental( i) y experimental( mean))( ypredicted( i) ypredicted( mean)) ( y experimental( i) y experimental( mean))^2 i= ( ypredicted( i) ypredicted( mean))^2 4 Results and Discussions 4. A Model Development for MISO (Multi Input Single Output) System: While the training set was utilized for the Error back propagation based iterative updation of the network weights, the test set was used for simultaneously monitoring the generalization ability of the multilayer perceptron (MLP) model. The MLP architecture comprised nine input ( = 9) and one output (K = ) nodes. For developing an optimal MLP model, its structural parameter, namely the number of hidden nodes (L) was varied systematically. For choosing an overall optimal network model, the criterion used was least AARE for the test set. The optimal MLP model that satisfied this criterion has thirty hidden nodes, tan sigmoid activation function at input and tan sigmoid activation function at output nodes. The average error (AARE) for training and test set is calculated as 0.04% and 0.042% and corresponding cross correlation co-efficient (R ) calculated as 0.84 and 0.83 respectively. The low and comparable training and test error AARE values indicate good prediction and generalization ability of the trained network model. Good prediction and generalization performance of the model is also evident from the high and comparable R values corresponding to both the outputs of training and test sets. Figure 4 depicts a comparison of the outputs as predicted by the MLP model and their target values. Considering the fact that all the input output data are from real plant with their inherent noise, the very low prediction error can be considered as an excellent A model. Once developed, this A model can be used to quantitatively predict the effects of all input parameters on the MEG product UV transmittance. 4.2 Actual Plant vs v Predicted UV in Plant: P To validate the reliability of model,actual plant data were taken from DCS at different plant load at different point of time and actual lab measured UV was compared with the model predicted UV. Fig: 4 depict the actual versus the predicted UV. 8

10 4.3 DE-based optimization of the A model After development of successful A model of glycol column, next step is to find out the best set of operating conditions which lead to maximum UV. DE based hybrid model was run and optimum parameters were evaluated ( within their permissible operating limit). Fig.5 depicts the actual versus the optimum UV. From figure 5 it is clear that by making a small change in the nine input parameters, the to 2% rise in UV can be made. The program was made online where it gives the operator what should be the nine input parameters at different time to maximized the UV in real time basis. After verifying all the calculations, the optimum input parameters were maintained in actual plant and benefit was found exactly same as calculated. This ensures the validation and accuracy of this calculation. Fig: 4 Actual Vs Predicted UV Fig: 5 Actual Vs s optimum UV 5. Conclusion In this paper, process modeling and optimization strategies integrating artificial neural networks with the differential evolution have been employed for modeling and optimization of commercial ethylene glycol product column. In the strategy, a process model is developed using an A method following which the input space of that model is optimized using DEs such that the process performance is maximized. The major advantage of the A-DE strategy is that modeling and optimization can be conducted exclusively from the historic process data wherein the detailed knowledge of process phenomenology (reaction mechanism, kinetics etc.) is not required. Using A-DE strategy, a number of sets of optimized operating conditions leading to maximized product UV was obtained. The optimized solutions when verified in actual plant resulted in a significant improvement in the product UV. References. Babu, B.V & Sastry, K.K..(999). Estimation of heat transfer parameters in a trickle-bed reactor using differential evolution and orthogonal collocation, Comp. Chem. Engg. 23, Tambe S. S., Kulkarni B. D. and Deshpande P. B. (996), Elements of Artificial eural etworks with selected applications in Chemical Engineering, and Chemical & Biological Sciences, Simulations & Advanced Controls, Louisville, KY. 9