Adaptive system identification of industrial ethylene splitter: A comparison of subspace identification and artificial neural networks

https://doi.org/10.1016/j.compchemeng.2021.107240Get rights and content

Highlights

  • Modeling an industrial ethylene splitter in Aspen Dynamics and validate the model.

  • Compare neural network system identification methods to subspace identification.

  • Develop an online model adaptation scheme to improve model prediction capabilities.

  • Adapt the system identification methods to simulated and real plant data.

Abstract

The manuscript considers the problem of data-driven modeling of an ethylene splitter (from an industrial plant). The process presently operates with end composition controllers that does not work well during process transition. The objective of the present work is to investigate the use of different data-driven techniques such as subspace identification and neural network-based methods for the purpose of developing a dynamic data-driven model. To this end, first an ethylene splitter simulation model is built that replicates industrial operation. The ability of the simulation model to capture the key traits of the process dynamics are first established by comparing it with data from the plant operation. The simulation model is subsequently utilized to work as a test bed for future control purposes and to serve as an additional test of the modeling approaches. An online model adaptation scheme is developed to improve the model's prediction capabilities under new operation patterns.

Introduction

Ethylene (C2H4) is one of the most versatile and widely used petrochemicals in the world today and is primarily being used for the manufacturing of polyethylene. The separation of ethylene from ethane by the C2 splitter is normally the final step in the production of ethylene, where the final products are primarily ethylene in the top stream and ethane in the bottom stream. The top ethylene product is then sold, while the bottom ethane is recycled back to upstream processes. Separation of ethylene from ethane is one of the most energy intensive separations, which uses large distillation column with over 100 trays due to the small difference in relative volatilities of ethane and ethylene (Salerno et al., 2011). Another contributor to the high energy consumption is that the ethylene splitter is commonly operated at high-pressure, utilizing closed-cycle propylene refrigeration. Optimization of the C2 splitter process is critical to meet desired ethylene product purity while maximization the ethylene-ethane separation and the hydraulic processing capacity of the tower. Continued push towards operational efficiency of the ethylene splitter process, in turn, continues to motivate advanced modeling and control efforts in this direction. Of particular interest are modeling approaches that can readily adapt the model to new operating conditions as new data becomes available.

The complexity of C2 splitter operation stems from many reasons: (1) high interaction between the top and bottom purities, (2) small difference in relative volatilities of ethane and ethylene, (3) slow dynamics of the column, (4) nonlinear dynamic behavior which changes in the plant conditions (such as pressure and temperature), and (5) continuous change in feed flow and composition. C2 splitter operates with limited capacity and high feed flow which can result in exceeding tower capacity causing tower flooding. The ethylene splitter can be modeled using first-principle models or data-driven models. The advantage of first-principle models is that they can work in a relatively broad range of operations, but such models are expensive, hard to develop, and harder to maintain. Therefore, industrial practice has sought to pursue data-driven models that are relatively easier to develop and maintain (Zhao et al., 2001). These challenges make the problem of building an accurate model for C2 splitter a hard task to accomplish (Kolmetz and Sari, 2014). The critical metric in evaluating the usefulness of such data driven models is model validity and extrapolation capabilities, especially when being used for the purpose of control and optimization (and not just process monitoring).

In the direction of utilizing first principle model for ethylene splitter, Salerno et al. (Salerno et al., 2011) developed a rigorous first principle model in Aspen Plus for ethylene splitter. Borralho (Borralho, 2013) developed a first principle model for the ethylene plant in gPROMS. Yan (Yan, 2000) developed a simplified first principle ethylene plant model which includes a thermal cracking section, a separation system, and an integrated refrigeration. Wang (Wang et al., 2014) developed a rigorous dynamic simulation for the startup operation of C2 splitter. Choe and Luyben (Choe and Luyben, 1987) developed rigorous dynamic models of C2 splitter that considers the effect of vapor holdup in high pressure columns, which is normally ignored. Eliceche et al. (Eliceche et al., 1995) studied ethylene plant units’ capacities for bottlenecks in the furnaces and ethylene splitter sections caused by changes in the feed flow rate and composition. All these contributions, however, used only simulations and have not validated their results against plant operation.

Friedman (Friedman, 1999) reviewed the existing contributions in modeling ethylene plant units using various data driven models, including subspace identification and artificial neural network-based approaches. Huang et al. (Huang et al., 2005) proposed a closed-loop subspace identification approach through an orthogonal projection. Several examples exist of the application of subspace identification methods for data driven modeling of distillation columns using simulation data (Meidanshahi et al., 2017; Castaño et al., 2011; Meenakshi et al., 2013). Kanthasamy et al. (Kanthasamy et al., 2014) developed a nonlinear system identification model for pilot plant distillation column based on Hammerstein model. The model consists of a nonlinear static element followed by a linear dynamic model. The data generated from their first principle model was first validated using experimental data before being used as the process model in the Hammerstein model parameter estimation. Norquay et al. (Norquay et al., 1999) developed a nonlinear system identification model for an industrial ethylene splitter based on Wiener model, which consists of a linear dynamic element followed by a nonlinear static element. In their work, they used plant data with only two outputs and three inputs along with generated simulation to develop the Wiener model and compared it to real plant data.

Artificial neural networks (ANN) models have been used by many researchers in the context of model identification of a distillation column. Many works have utilized simulation data to model distillation column using feed forward neural network (Brizuela et al., 1996; Singh et al., 2005; Ochoa-Estopier et al., 2013), recurrent neural network (MacMurray and Himmelblau, 1993; Pan et al., 2001) and nonlinear autoregressive with exogenous inputs (NARX) network architecture (Jaleel and Aparna, 2015; Jaleel and Aparna, 2019). Relatively fewer contributions model industrial distillation columns using real process data. Savkovic-Stevanovic (Savkovic-Stevanovic, 1996) used plant data to develop a feed forward neural network model for an industrial distillation column. Singh et al. (Singh et al., 2013) used laboratory data to model a 9-tray binary distillation column available in the laboratory using both, feed forward neural network and recurrent neural network. Abdullah et al. (Abdullah et al., 2009) proposed a feed forward neural network to predict the top and bottom product composition of a pilot plant distillation column using simulation data and validation against plant data. ANN based Model Predictive Controller (MPC) has also been used in other application (Sadeghassadi et al., 2018; Zhang et al., 2019). In summary, while many contributions have utilized ANN models based on either simulation data or real data, limited work has been carried out to validate the model with real data from industrial distillation columns and to compare ANN based approaches against alternatives such as subspace identification.

In practice, process operation changes over time due to internal and external conditions which can cause deterioration in model predictions. Therefore, the need to continuously updating process models, via adaptive modeling algorithm, as time evolves might be necessary to sustain the model predictions accuracy. Alanqar et al. (Alanqar et al., 2017) proposed an error-triggered on-line model identification strategy for linear state-space model to obtain more accurate state predictions for nonlinear process systems. The error-triggering was conducted by a moving horizon error detector that quantifies the relative prediction error within its horizon and triggers model re-identification using recent operations when the prediction error exceeds a threshold. Wu et al. (Wu et al., 2019) proposed a machine learning-based predictive control scheme that utilizes an online update of the Recurrent Neural Network RNN models to capture process nonlinear dynamics in the presence of model uncertainty.

Motivated by these considerations, in the present manuscript, the ethylene splitter, a distillation column with a large number of inputs and output is modelled using different system identification methods. The system identification methods studied in this work are subspace identification, NARX neural network, and nonlinear RNN. This is achieved in two steps. First, a simulation model in Aspen Dynamics is developed to work as a test bed for future control purposes and to serve as an additional test of the modeling approaches. The Aspen simulator is developed for testing our implemented control strategy, since it cannot be demonstrated directly using the real plant. Subsequently, various data driven models are developed and tested against both the simulation and the plant data. Finally, two online modelling schemes are developed for the purpose of continuously updating the models with new available data to improve their predictions capabilities. The two adaptive modeling algorithms are introduced to keep the dynamic model up to date with the most recent operations and these algorithms are adapted for our three system identification methods. The rest of the paper is organized as follows. Section 2 presents an overview of the ethylene plant,the C2 splitter, and the system identification methods. Section 3 presents the developed Aspen dynamics-based simulation model and compares the simulation results against the plant operation. Section 4 presents the system identification results. Section 5 presents the two adaptive strategies for three different system identification methods that allow more recent data to be incorporated into the training approaches. Section 6 demonstrates the importance of using an adaptive modeling scheme when modeling the ethylene splitter for the three system identification methods. Finally, concluding remarks are presented in section 7.

Section snippets

Preliminaries

In this section, a detailed description of the C2 splitter is provided, followed by a brief review of the different system identification methods utilized.

Modeling ethylene splitter

As noted earlier, the first step was to create a dynamic model for testing of the modeling procedure and as a future test bed for feedback control purposes. Since the tower has been simulated as an isolated system, the dynamic model needed adjustment to replicate the industrial process shown previously in Fig. 1 due to two issues. First, the model does not consider the effect of the propylene stream through the heat exchanger used to preheat the feed. Second, the tower operates with pressure

Data driven modeling of the ethylene splitter

In this section, we model the ethylene splitter to evaluate the different system identification methods for three different cases: (1) simulated data with 9 inputs shown in Table 1 (Sim9), (2) simulated data with only 4 inputs (feed flow, feed temperature, reboiler duty, and reflux flow) (Sim4), and (3) plant data with the same 4 inputs as case 2 (Plant4). The Sim9 case study uses all of the available tower inputs to model the ethylene splitter, including inputs that are not measured in the

Online model updating scheme

In the previous section, we illustrated our different system identification methods for the ethylene splitter and analyzed the influence of different hyperparameters on each model's performance. Finally, we evaluated our developed model on new data (testing dataset). In this section, we propose different online system identification algorithms to improve the model predictions capability and evaluate it. One of the key things to consider when implementing system identification online is model

Ethylene splitter online modeling scheme results

To demonstrate the online modeling schemes introduced in the previous section, we use the same training and testing datasets used earlier for the three case studies. In this section, we consider 1,500 training samples (n) for case studies Sim9, Sim4, and Plant4. Also, we consider 500 testing samples (m) for testing all case studies. The 500 testing samples are predicted recursively 20 samples (Δm) at a time with an updated model. The 20 samples which correspond to a duration of 3 hours and 20

Conclusions

In this paper, three different system identification methods (subspace identification, NARX neural network, and RNN) have been developed for modeling industrial ethylene splitter. The system identification methods have been adapted to fit simulated and real plant data to compare and show their capability at developing dynamic model for both data. The results show that linear subspace identification method can provide the best predictions in general for the dynamic systems due to their

CRediT authorship contribution statement

Mahir Jalanko: Conceptualization, Methodology, Software, Writing - original draft, Writing - review & editing. Yoel Sanchez: Conceptualization, Methodology, Investigation, Writing - review & editing. Vladimir Mahalec: Conceptualization, Methodology, Supervision, Visualization, Writing - review & editing. Prashant Mhaskar: Conceptualization, Methodology, Supervision, Writing - review & editing.

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Acknowledgments

This work was supported by NOVA Chemicals Corporation, Ontario Graduate Students, and McMaster Advanced Control Consortium is greatly acknowledged

References (42)

  • V. Singh et al.

    ANN based estimator for distillation—inferential control

    Chem. Eng. Process.: Process Intensif.

    (2005)
  • H. Zhao et al.

    A nonlinear industrial model predictive controller using integrated PLS and neural net state-space model

    Control Eng. Pract.

    (2001)
  • Abedi, A.A., 2007. Economical analysis of a new gas to ethylene technology (Doctoral dissertation, Texas A&M...
  • A. Alanqar et al.

    Error-triggered on-line model identification for model-based feedback control

    AIChE J.

    (2017)
  • F.J.O. Borralho

    Detailed Modelling and Optimisation of an Ethylene Plant (Doctoral dissertation

    (2013)
  • E. Brizuela et al.

    Predictive control of a multi-component distillation column based on neural networks E. Brizuela

  • J.E. Castaño et al.

    Model identification for control of a distillation column

  • Y.S. Choe et al.

    Rigorous dynamic models of distillation columns

    Ind. Eng. Chem. Res.

    (1987)
  • B. Corbett et al.

    Subspace identification for data-driven modeling and quality control of batch processes

    AIChE J.

    (2016)
  • Friedman, Y.Z. (1999). Advanced control of ethylene plants: what works, what doesn't, and why. Ethylene Producers’...
  • A. Garg et al.

    Subspace identification-based modeling and control of batch particulate processes

    Ind. Eng. Chem. Res.

    (2017)
  • Cited by (21)

    • Optimization and control of extractive distillation for formic acid-water separation with maximum-boiling azeotrope

      2023, Computers and Chemical Engineering
      Citation Excerpt :

      The control performance of model predictive controller (MPC) was further enhanced by modifying the continuous linear time-invariant (CLTI) model into an identified model (Feng et al., 2019). Besides, nonlinear model predictive control (NMPC) based on neural network model and transient wave propagation model were developed to control complex ethylene splitter (Jalanko et al., 2021a, 2021b) and air separation unit (Schulze et al., 2021). Therefore, MPC is expected to achieve better control performance for complex process such as azeotrope separation in present work since it considers the interactive effects of all manipulated variables to all controlled variables.

    • Adaptive non-linear soft sensor for quality monitoring in refineries using Just-in-Time Learning—Generalized regression neural network approach

      2022, Applied Soft Computing
      Citation Excerpt :

      The training of NARX neural network is similar to feed forward neural networks used to predict one step or multi-step ahead prediction of output variables using past instants of input and output along with current instant of incoming query objects. There are two steps in NARX for estimation: Open loop structure and Closed loop structure [38]. Pictorial representation of NARX model is shown in Fig. 3.

    View all citing articles on Scopus
    View full text