468067 Development of a High Fidelity and Reduced Order Data-Driven Model of Hydrogen Plant

Tuesday, November 15, 2016: 2:24 PM
Carmel II (Hotel Nikko San Francisco)
Abhinav Garg1, Brandon Corbett2, Prashant Mhaskar2, Gangshi Hu3 and Jesus Flores-Cerrillo4, (1)Department of Chemical Engineering, McMaster University, Hamilton, ON, Canada, (2)Chemical Engineering, McMaster University, Hamilton, ON, Canada, (3)Praxair, (4)Praxair, Tonawanda, NY

Hydrogen is one of the most indispensable chemical components extensively used in petroleum and chemical industries. An economic way to commercially produce hydrogen is to utilize steam methane reforming [1]. The process consists of a number of material and heat flows, resulting in an intricate network of material and energy. Natural gas (NG) and superheated steam are fed to a chemical reactor called reformer, which consists of catalyst tubes filled with nickel reforming catalyst. Here, majority of the hydrogen is produced. The gas out of the reformer is then processed through another reactor to further produce hydrogen. Finally, the hydrogen flow is purified in a pressure swing adsorber (PSA), where high purity hydrogen is produced. The reformer exit temperature is an important process variable for this process and is expected to be kept at a desired level by heating the reformer. This heat is provided by burning the off-gas from the PSA and NG fuel stream. A fan is used to supply air to the burners and another one draws the combustion products, termed as flue gas, out of the reformer box. This process presents various constraints and challenges during the start-up as well as nominal operation of the plant such as, the reformer exit temperature should be maintained at a desirable level. Further, the firebox pressure should not breach its lower and upper limits for safety. If the pressure is too low, the fire can be extinguished. If it is too high, it may impose safety hazards to facility and personnel [2].

A typical startup procedure consists of series of events that take place over a period of time based on a predefined recipe. Startup can broadly be categorized as cold startup and warm startup based on the initial state of the plant from which the plant is required to be driven to nominal operation. The focus of this work will be on cold startup where the plant is started from completely shutdown conditions. Typical steps involve ramping the natural gas fuel to reach a sufficient reformer exit temperature, circulation of natural gas feed and steam, regulation of nitrogen flow in the reactor tubes. The startup also involves making discrete decisions such as starting up the PSA, recirculating the tail gas from the PSA and ramping up to full capacity in several stages. Thus, the startup time is not fixed and varies based on the decisions taken at various stages. Existing startup recipes have been built heuristically from past successful startups, with the possibility of optimizing the startup procedure remaining unexplored.

One key recognition that we make is that due to the nature of startup/shutdown phase, they can be treated as a batch like operation with variable lengths, where the objective is to drive the process from one state to the other. For modeling and control of batch processes, a variety of a data-driven modeling techniques exist. One of the most widely used approach is partial least squares (PLS), which models the process in a projected latent space [3]. These models are essentially time-varying linear models, linearized around mean past trajectories, and therefore require the batches to be of same length, or to recognize an appropriate alignment variable. To account for these limitations, a multi-model approach was proposed in [4]. These models were based on the 'current measurements' of the process instead of the 'time'. These developments were followed by contributions in the area of integration of these data-driven models with the advanced control formulations [4-5]. More recently a subspace identification based batch control approach was proposed in [6] where a LTI state-space model of the batch process is estimated, and does not require the training batches to be of equal length. The existing formulations have, however, not been applied to startup/shutdown operations in general, and for standard operating procedure (SOP) synthesis/optimization problem for the hydrogen plant in particular.

Motivated by these considerations, in this work we address the problem of determining a data driven model for the startup of a hydrogen plant. To this end, first a high fidelity test bed model of the entire plant is developed in UniSim, capable of simulating the startup and shutdown phase, with appropriate adaptation of the plant SOP. Several simulated startups are performed to generate the training data. Then, an LTI data-driven model (higher order) of the process using subspace identification based method [6-7] is estimated. Simulation results are presented that illustrate the prediction capabilities of the identified model and its suitability for deployment in an optimization framework.

Note that the development of a good model that can handle variable batch times paves the way for application of optimization algorithms that can improve not just the process economics along the startup/shutdown but also the operating time of the startup/shutdown procedure. From an optimization standpoint, the problem is quite different compared to existing application of data driven models for control of lumped parameter (such as in [6]) or distributed parameter process [7], where the objective was to reach a certain quality requirement in a given, predetermined batch length. The optimization problem for the present application is an optimal control problem where the objective is to reach the desired terminal state in minimum time, while adhering to various constraints, and also making discrete decisions along the way. The data driven model developed in the present work will subsequently be utilized within such an optimization framework to synthesize the startup SOP.

[1] JL Lynn and BH Bory. Kirk-othmer encyclopedia of chemical technology, 2000.

[2] Miao Du, Prashant Mhaskar, Yu Zhu, and Jesus Flores-Cerrillo (2014). Safe-Parking of a Hydrogen Production Unit. Industrial & Engineering Chemistry Research, 53 (19), 8147-8154.

[3] J. Flores-Cerrillo, J. F. MacGregor, Control of particle size distributions in emulsion semibatch polymerization using mid-course correction policies, Industrial and Engineering Chemistry Research 41 (7) (2002) 1805-1814.

[4] S. Aumi, B. Corbett, P. Mhaskar, and T. Clarke-Pringle. Data-based modeling and control of nylon-6, 6 batch polymerization. IEEE Transactions on Control Systems Technology, 21(1):94-106, Jan 2013.

[5] S. Aumi, P. Mhaskar, Integrating Data-Based Modeling and Nonlinear Control Tools For Batch Process Control, AIChE Journal 58 (2012) 2105-2119.

[6] B. Corbett, P. Mhaskar, Subspace identification for data-driven modelling and quality control of batch processes, AIChE Journal (2016) n/a-n/a.

[7] Abhinav Garg, P. Mhaskar, Data-Driven Model Predictive Control of Crystal Size Distribution in a Seeded Batch Crystallization Process, Chemical Engineering Science (2016) Submitted.


Extended Abstract: File Not Uploaded
See more of this Session: Industrial Applications in Design and Operations
See more of this Group/Topical: Computing and Systems Technology Division