High Level Optimization Model for the Retrofit Planning of Process Networks

The retrofit design of a network of processes over several time periods is addressed in this paper. A strategy is proposed that consists of a high level to analyze the entire network, and a low level to analyze a specific process flowsheet in detail. A methodology is presented for the high level to model process flowsheets and retrofit modifications using a multiperiod generalized disjunctive programming (GDP) model. This problem is reformulated as a mixed-integer linear program (MILP) using the convex hull formulation. Two examples that illustrate the proposed model are presented. The results show that the proposed GDP model provides a significant benefit over the existing network without retrofit, and provides a clear advantage over intuitively choosing modifications based on heuristics. To illustrate the performance benefits of using the convex hull formulation, the problem is also modeled as an MILP with big-M constraints. Introduction The area of process synthesis has mainly concentrated on integrating subsystems and flowsheets for the design of new processes, or “grassroots” design. Many algorithms and techniques have been developed to synthesize and optimize processes as different areas of process synthesis become better understood. The less thoroughly explored area of retrofit synthesis, however, addresses one of the most prevalent problems in the chemical processing industry, which is the evaluation and redesign of existing plants. Processes can be retrofitted to achieve such goals as increasing throughput, reducing energy consumption, improving yields, and reducing waste generation. Retrofit designs may involve relatively small modifications (e.g. addition of equipment), or essentially tearing down the existing plant and replacing it by a new one. The retrofit problem is difficult due to the many constraints of a preexisting operation such as layout, available space, piping, and operating conditions. It is the goal of this work to address retrofit design of networks of chemical processes. This is a problem that has received little attention in the literature. Work in retrofit design has been limited due to the difficulties described above and due to the many modification possibilities, which causes the problem to grow greatly in size. For instance, Grossmann et al. show that for retrofit of a distillation sequence separating a mixture of N components, the number of tasks considered is N 1 times that for the grassroots design. The combinatorial nature of retrofit synthesis is the reason that research in the area to date has focused primarily on modifying a particular subsystem or equipment type. For a general review of retrofit issues see Grossmann et al. As with the work in process synthesis, most literature for retrofit deals with Heat Exchanger Networks (HENs). Recent work on HENs retrofit includes procedures proposed by Kralj et al. who use a stepwise simultaneous superstructural approach, Bochenek and Jezowski who use an adaptive random search method, and Briones and Kokossis who applied a mathematical programming and pinch analysis decomposition scheme and later used retrofit Hypertargets in a conceptual programming approach. Other subsystem work has been done by Papageorgaki and Reklaitis and subsequently by Georgiadis et al. who proposed optimal retrofit procedures for multi-purpose batch chemical plants. Seo et al. developed optimization techniques for the redesign of crude distillation columns. And Fraser and Hallale proposed a method for the retrofit of mass exchange networks using the pinch technology that was developed for HENs. Works that deal with retrofitting subsystems while also taking into account the retrofit of effected process networks include a method by Linhoff and Eastwood that uses pinch technology for overall site optimization, and an approach by Zhang and Zhu that uses simultaneous optimization of HENs and process changes. Strategies to solve retrofit problems at the level of an individual flowsheet have been proposed by Grossmann et al. who give a general outline for dealing with process retrofits, and by Fisher et al., who present a systematic procedure for developing and screening retrofit opportunities. The method by Fisher et al., based on Douglas’ hierarchy for design, is perhaps the most comprehensive to date, but is not systematic for predicting the retrofit changes. In spite of the work described above, there still exists no systematic strategy that unifies the retrofit design of all subsystems in an existing process or network of processes. Companies in the chemical processing industry and related industries often have several site locations worldwide. Each of these sites may contain multiple processing plants which themselves may contain a series of different processes. These types of large-scale operations make the task of considering where to make retrofit improvements a difficult one. The application of an optimization model to this problem would be quite useful. However, the formulation of detailed models for each of the processes in a plant network is a cumbersome task involving the collection of many types of data. In fact, the data collection step alone may be too time consuming to make process modeling worthwhile. Therefore, we aim to provide a screening tool where the optimization model is formulated from a limited amount of process data. The focus of this paper will be at the level of retrofitting a network consisting of several interconnected processes. A methodology is proposed that relies on a multiperiod generalized disjunctive programming (GDP) model that is reformulated as a mixedinteger linear program (MILP). Problem Statement The specific problem is as follows. An existing network of chemical processes is given, for which each process can possibly be retrofitted for improvements such as higher yield, increased capacity, and reduced energy consumption. Given limited capital investment to make process improvements and cost estimations over a specified time horizon, the problem consists of identifying those modifications that yield the highest economic improvement in terms of Economic Potential (EP) . We define EP as the income from product sales minus the cost of raw materials, energy, and process modifications. Methodology We propose to address the retrofit problem using a hierarchical approach and mathematical programming tools. Unlike the hierarchy of five decision levels proposed by Douglas for process synthesis, our method uses two levels: a high level for simultaneously analyzing the entire network, and a low level for analyzing a specific process flowsheet in detail. The focus of this paper is on the development of the high level model. The nature of our retrofit problem requires the high level model to be fairly detailed in order to capture all the design and operation parameters most affected by process modifications. As an initial step, we assume that the high level model is linear. This is done to ease the computational effort in solving the large and complex proposed problem. Mass and energy balances are represented with linear equations, meaning that operating conditions are assumed to be fixed with constant temperatures and pressures for all process streams. Our aim is to model the proposed retrofit problem by developing a multi-period Generalized Disjunctive Programming (GDP) model that is reformulated into a multi-period MILP using convex hull transformations for each disjunction. This approach is in contrast to the use of conventional MILP formulations that rely on big-M constraints. For the high level model, a linear programming (LP) model is developed from the process flowsheet using the procedure outlined below. An aggregated model is used to predict bounds for best process performance information independent of detailed design modifications. To construct the model for each process in the network (Figure 1) we proceed as follows: 1. Represent each flowsheet using only essential equipment, combining units when possible. 2. Set the operating conditions, fixing all stream temperatures and pressures. 3. Establish linear mass and energy balances that account for stoichiometry, yield, energy irreversibilities, etc. 4. Use balances in step 3 to formulate an LP model of the process that will determine all process flow rates. 5. Identify potential improvements and estimate the cost to implement them. It should be noted that the first four steps are described in detail in Biegler et al.. Step 5 makes use of a variety of tools and techniques in process synthesis (e.g. pinch analysis, attainable region). This step represents the major challenge in the proposed methodology. We assume in this paper that major modifications considered are typically for increasing production capacity and/or conversion, and for improving energy recovery. High Level GDP Model The proposed model is next set up to identify potential improvements that are to be realized through retrofit projects over T time periods and with limited capital in each of these periods. To model the modifications we will use Generalized Disjunctive Node n Reactor Condenser Distillation Process i Figure 1. Sample Process Network and Process Flowsheet Programming, a logic-based representation of mixed integer problems where constraints associated with discrete variables are represented through disjunctions. The objective function to be maximized is the economic potential over a planning horizon of T time periods. Costs for reactants, energy and modifications, and profits from product sales are included in the EP. Disjunctions are used to model changes in operating modes and design modifications. Operating modes i M m ∈ for each process i correspond to either existing operating conditions( ) i M m ∈ 0 , or to new conditions ( ) 0 , m m M m i ≠ ∈ that result from a retrofit project (e.g. higher yield, higher throughput, lower energy). The Boolean variables t im Y represent the selection of operating mode m for process i in each time period t, while t im W represents the decision to make design change m for process i in period t. Note that t im W can be regarded as the Boolean variable for project selection, while t im Y can be regarded as the Boolean variable that defines the operation as a function of the decisions t im W . i M is the set of operating modes m for each process i, with the first element corresponding to no changes to the existing process i. The set of process balances that correspond to each operating mode m in each time period t can be written as: T t M m P i b x A i t im t i t im ... 1 , , = ∈ ∈ ≤ (1) The cost t i c for modifying process i in time period t is given by: T t M m P i FC c i t im t i ... 1 , , = ∈ ∈ = (2) where t im FC is the fixed cost associated with design change m. t LC specifies the limit on funds available to invest in retrofit projects in time period t. Funds not invested in a given time period do not accrue for use in subsequent time periods. The continuous variables t i x and t n x correspond to flow rates associated with process P i ∈ at time t and node N n ∈ (mixers and splitters) at time t, respectively. Chemical and energy prices are given by p. The GDP model (P1) with logic constraints is as follows: