Estimate of process compositions and plantwide control from multiple secondary measurements using artificial neural networks
Introduction
One of the main difficulties in chemical process control is getting reliable and accurate measurement of product compositions, but the product compositions normally mean the product quality. Without the measurement of compositions, several good control schemes or theories cannot be normally implemented in chemical process industries. Although composition analyzers, like on-line gas chromatography (on-line GC), have been used in the process industries for a long time, they usually suffer from many shortcomings. The on-line GC usually needs a high investment. However, the reliability of the on-line GC is normally not so good, and it needs more cost and manpower for maintenance. Moreover, the most important drawback for using the on-line GC in chemical process control is that it possesses a very large time delay and thereby lowers the achievable control performance. The use of inferential variables to estimate process compositions in place of direct on-line measurements is usually desired for plant engineers. One of the most popular alternatives to on-line GCs in the control of continuous distillation column is using tray temperatures to infer the product compositions (Boyd, 1979, Yu & Luyben, 1984). Several researchers (Joseph, Brosilow, Howell & Kerr, 1976, Yu & Luyben, 1987, Mejdell & Skogestad, 1991a, Mejdell & Skogestad, 1991b, Piovoso & Kosanovich, 1994) have proposed various estimation techniques by using multiple temperatures to infer the product compositions in the distillation column. These techniques are normally based on rigorous models or statistical methods using steady-state plant data. One drawback for these methods is that they usually need a longer time, say 1 year, to collect the plant operation data. Recently, a dynamic estimator to infer the product compositions of a binary distillation column via temperature measurements using a nonlinear extended Kalman filter was also developed by Baratti, Bertucco, Da Rold, and Morbidelli (1995). From the above literature review, it seems that most previous studies in composition estimator were generally focused on a single unit, say distillation column. However, process systems are normally represented as sets of interconnected unit operations, and the column compositions for a typical process plant may also be influenced by the operation conditions of other units (e.g. reactor), and vice versa. As a result of more stringent environmental regulation and economic considerations, plantwide control recently has tended to become more important in today's chemical plants (Downs & Vogel, 1993). The single unit approach for composition estimator sometimes may not be satisfied by many chemical process industries, especially for those processes having recycles. Thus, to develop a composition estimator under plantwide consideration seems to be important.
On the other hand, artificial neural networks (ANN) have been successfully used for a number of chemical engineering applications. Astrom & McAvoy, 1992, Morris, Montague & Willis, 1994, Baughman & Liu, 1995 have presented overviews of the issues pertaining to the use of ANN for sensor data analysis, fault diagnosis, process modeling, identification and control. An ANN is composed of nets of nonlinear basis functions; it has the ability to evolve a good process model from experimental data and requires very little or no knowledge of first principles. It has the ability of learning and prediction for nonlinear models, and has therefore been used to identify the process dynamics nonparametrically by several authors. In addition, the ANN software that provides the inferential estimation for a process is also called a ‘virtual sensor’. Composition estimators using ANN for a batch distillation column via tray temperatures and flow rates have been recently developed by Zamprogna, Barolo, and Seborg (2001).
The primary object of this investigation, however, is to estimate the dynamic process compositions from other secondary measurements using the predicted capability of ANN under plantwide consideration. It assumes that the process plant has already installed on-line GCs, and the proposed estimator could be an aid of (or even in place of) on-line GCs in plant operations and control. In comparison with most of the previous estimation techniques (Joseph et al., 1976; Yu & Luyben, 1987, Mejdell & Skogestad, 1991a, Mejdell & Skogestad, 1991b, Piovoso & Kosanovich, 1994) using steady-state multiple tray-temperatures in a single column only, the proposed ANN model dynamically infers the process compositions using all possible measurements under plantwide consideration. The Tennessee Eastman (TE) plant (Downs & Vogel, 1993) with on-line GCs, which is typically a highly integrated and interconnected system, is employed for the investigation. A classic PI control scheme using the proposed ANN estimator in the TE plant is studied. Besides, a cascade configuration of model predictive control (MPC) technique using the proposed ANN estimator in the TE plant is also explored.
Section snippets
Process description
The TE plantwide industrial process developed by Downs and Vogel (1993) is a challenging problem in process control and related fields. Several researchers (McAvoy & Ye, 1994, Ricker, 1995, Ricker & Lee, 1995a, Ricker & Lee, 1995b, Banerjee & Arkun, 1995, Ricker, 1996, Luyben, 1996, Sriniwas & Arkun, 1997, McAvoy, 1999) have studied this process. A schematic diagram of the process is given in Fig. 1. The process involves five unit operations: a two-phase reactor, a condenser, a vapor/liquid
Composition estimator using artificial neural networks
The feedforward ANN structure as shown in Fig. 3 is composed of many interconnected neurons organized in successive layers and is employed in this study. The input layer is composed of fan-out units which do not perform any computation but simply distribute their inputs to all neurons in the next layer. The ‘bias’ as shown in Fig. 3 is considered to be similar to a neuron, and its weight is connected to a fixed input of +1. The last layer is the output layer. Between the input and output layers
Classic control using a composition estimator
Since the compositions XA,pg, XB,pg and XG,pd can be estimated from other secondary measurements by an ANN model, we try to use them as controlled variables based on Luyben (1996) control structure shown in Fig. 2. In order to eliminate inherent errors of the ANN estimator, a correction error for each component, as mentioned in the previous section, given by the plant laboratory every 2 h is also added. Then, the estimated composition is used instead of the on-line GC through the
Model predictive control using composition estimators
Early work on the TE process using nonlinear MPC was reported by Ricker and Lee (1995a). They presented a model of the TE process more rigorously with the help of lower order first principle model and on-line estimation. This model is a nonlinear mechanistic, state-variable formulation with 26 states, 10 manipulated variables and 23 outputs (Ricker & Lee, 1995b). Using a multivariable controller with eight controlled and manipulated variables, Ricker and Lee (1995a) have demonstrated that the
Conclusions
A dynamic approach to estimate process compositions from multiple secondary measurements under plantwide consideration using an ANN model has been investigated in this study. Recent publications in the literature concerning the composition estimator (Mejdell & Skogestad, 1991a, Mejdell & Skogestad, 1991b, Piovoso & Kosanovich, 1994) generally lie in the use of statistical methods, and they normally need a longer time to collect steady-state plant operating data. An alternative nonlinear
Acknowledgements
This work is supported by the National Science Council of Taiwan (R.O.C.) under the grant NSC85-2214-E-029-002.
References (29)
- et al.
Intelligent control
Journal of Process Control
(1992) - et al.
Control configuration design applied to the Tennessee Eastman plant-wide control problem
Computers and Chemical Engineering
(1995) - et al.
Development of a composition estimator for binary distillation columns. Application to a pilot plant
Chemical Engineering Science
(1995) - et al.
A plant-wide industrial process control problem
Computers and Chemical Engineering
(1993) - et al.
Modeling and control of a packed distillation column using artificial neural networks
Computers and Chemical Engineering
(1995) - et al.
Base control for the Tennessee Eastman problem
Computers and Chemical Engineering
(1994) - et al.
Process identification using neural networks
Computers and Chemical Engineering
(1992) Optimal steady-state operation of the Tennessee Eastman challenge process
Computers and Chemical Engineering
(1995)- et al.
Nonlinear model predictive control of the Tennessee Eastman challenge process
Computers and Chemical Engineering
(1995) - et al.
Nonlinear modeling and state estimation for the Tennessee Eastman challenge process
Computers and Chemical Engineering
(1995)
Composition estimations in a middle-vessel batch distillation column using artificial neural networks
Chemical Engineering Research and Design
Neural networks for bioprocessing and chemical engineering
Continuous-distillation column control
Multi-temps give better control
Hydrocarbon Processing
Cited by (30)
Pressure drop measurement for fine particulate slurry using optimization process
2022, Materials Today: ProceedingsCitation Excerpt :Numerous studies focused on different kinds of ANN models. Modeling of multiphase reactors [12], applications in process control [13–22], predictions from experiments in catalysis [23], to estimate the composition of a mixture [24], modeling of pump-mixer characteristics [25], chemical reactor modeling [26], experimental investigation of waste water treatment applications [27–31], and so on have all been efficaciously practiced with ANNs. In the existing investigation, an F.A. suspension was considered to predict the pressure drop using the ANN tool.
Artificial Intelligence techniques applied as estimator in chemical process systems - A literature survey
2015, Expert Systems with ApplicationsCitation Excerpt :They can be developed using software such as MATLAB and LabView and implemented on the specific process unit to predict unmeasured states such as concentration, temperature, heat flux, molecular weight and impurities. Other practical simulation software used for estimating states are PROCESS (Simulation Sciences code) (Himmelblau, 2008), SIAMOD (Siemens adaptive modelling of dynamic system) (Horn, 2001), NeuralWorks Professional II/(PLUS) (Yeh, Huang, & Huang, 2003) and dynamic simulator (Dynafrag) (Du, del Villar, & Thibault, 1997). Before AI-based estimators were introduced, researchers have been using the conventional observers in chemical process systems, for example the Luenberger observer, Extended Kalman Filter (EKF), sliding mode observer and observers.
Benchmark problems for nonlinear system identification and control using Soft Computing methods: Need and overview
2014, Applied Soft Computing JournalCitation Excerpt :Set-point change performance is assessed by the required transition time. In [99], a recurrent/dynamical neural network is used to predict the compositions of species mixtures instead of using on-line gas chromatographs. The predictions are used to improve conventional PI control and for model-predictive control.
Predicting effective viscosity of magnetite ore slurries by using artificial neural network
2012, Powder TechnologyCitation Excerpt :The ANN model can learn from examples, incorporate a large number of variables and provide a useful measure to learn from any new information, which was not previously used to train them [14–16]. The ANNs have been successfully applied to problems in various fields of chemical engineering, such as modeling of multiphase reactors [17], applications in process control [18–22], predictions from experiments in catalysis [23], to estimate composition of a mixture [24], modeling of pump-mixer characteristics [25], modeling behavior of membrane separations [26], chemical reactor modeling [27], experimental investigation of applications of waste water treatment [28] etc. In this paper, we measured and studied relations such as viscosity-solids concentration data and this was later tested for six empirical models, which were obtained as standard results from literature review.
Mixture probabilistic PCR model for soft sensing of multimode processes
2011, Chemometrics and Intelligent Laboratory SystemsCitation Excerpt :Thirdly, many data mining and modeling methods have been intensively researched, which can give strong technique supports for successful development of soft sensors. To our knowledge, most data-based soft sensors are based on principal component regression (PCR), partial least squares (PLS), artificial neural networks (ANN), support vector regression (SVR), and etc. [1–7,14–23]. According to the recent review paper on the data-based soft sensor [7], PCR is the most widely used method for soft sensor modeling, a total of 23% of soft sensor models have been constructed by the PCR method.
Soft sensor based composition estimation and controller design for an ideal reactive distillation column
2011, ISA TransactionsCitation Excerpt :Bahar et al. [12] demonstrated that an artificial neural network can be successfully designed for the estimation of composition measurements from temperature measurements throughout the column in a reactive batch distillation column. However, only very few papers have discussed the application of neural networks for the composition estimation in a continuous reactive distillation column [13,14]. The main contributions of the paper are to demonstrate the effectiveness of a recurrent neural network to estimate the unknown composition of a reactive distillation column based on available temperature measurements.