January 2019

Special Focus: Process Optimization

Improving resource efficiency of industrial processes with TOP-REF methodology—Part 1

The processing industry consumes significant amounts of resources, such as materials and energy, to produce valuable goods for society.

The processing industry consumes significant amounts of resources, such as materials and energy, to produce valuable goods for society. This results in a large footprint on the environment. Resource efficiency is a key factor for industry to decrease its footprint, achieve its sustainability goals and contribute to an efficient use of resources.

The European TOP-REF research consortium developed and applied a methodology to improve the resource efficiency of industrial production processes. A newly developed key resource indicator (KRI) that enables the consideration of material and energy streams in a single indicator—the resource exergy indicator (REI)—was established. Part 1 describes the TOP-REF methodology. Part 2, which will be featured in the February issue of Hydrocarbon Processing, presents its application to three industrial production plants from different industrial sectors: a granular NPK fertilizer production plant, a steam cracker plant, and the crude distillation and fractionation section of a crude oil refinery. In addition, this work aims at sharing the authors’ experiences and findings, while developing and applying the TOP-REF methodology with an outlook for further developments.

The five KRIs

A KRI indicates the resource efficiency of a process. Since resource is a manifold term that can be related to different types as material or energy, several KRIs were identified that apply relevant, accepted, credible, easy-to-measure and robust (RACER) criteria.¹

The resulting five KRIs were:

1. Unit material cost (UMC)
2. Direct primary energy consumption (DPEC)
3. Gross water use (GWU)
4. Net water use (NWU)
5. REI (TABLE 1).

The first four KRIs are more conventional indicators that are already applied in industry, whereas the fifth KRI—REI—is a novel and aggregated indicator. With the REI, it is possible to combine energy and material streams into a single unit based on exergy, and to merge the conventional indicators to one. Exemplary calculations of the KRIs can be found in the standard CWA 17185:2017.

TOP-REF methodology

The methodology of the TOP-REF project is to improve the resource efficiency of industrial processes. FIG. 1 provides an overview of the different steps within this methodology.

FIG. 1. Overview of the TOP-REF methodology.

The basis of the methodology is formed by process models that are detailed models based on physical and chemical phenomena. Such detailed process models enable the investigation of operating points away from the typical range of operation. However, the necessary computational time to run these detailed process models is high. This proves to be limiting when using these models for a subsequent global sensitivity analysis and optimization. In addition, it might be necessary to divide the process model generation into smaller subsystems due to various reasons. For example, this need can occur due to the complexity of the system or a requirement for specialized software to model subsystems of a special field of expertise.

To enable the utilization of the process models for a global sensitivity analysis and a subsequent optimization, a surrogate model is generated. This surrogate model is a simpler, computationally faster and more robust mathematical approximation of the detailed process model. Additionally, the surrogate model generation is used to couple the independently developed detailed process models of subsystems. Over the course of the project, for the case studies of the steam cracker and the crude oil refinery, for example, surrogate model generation was needed to couple respective process and utility section process models.

Based on the surrogate models, a global sensitivity analysis was performed to identify the process parameters (PPs) with the largest influence on the defined KRIs. These identified critical process parameters (CPPs) were then optimized to improve the KRIs of the considered processes in the following step.

Prior to the application of the optimal set of CPPs to the real production plant, the related monitoring and control variables to the optimized CPPs were identified. Within the demonstration, the set of optimal CPPs was applied to the real production plants. This demonstration permitted the testing of the theoretically developed and applied methodology to the real world. Therefore, the theoretically determined improvement in resource efficiency was compared to the real plant results.

Process models

The process models form the basis for the TOP-REF methodology and are of utmost importance. To enable the investigation of operating points away from the typical range of operation, the processes under consideration needed to be modeled in detail, based on physical and chemical phenomena. The first step in the model generation was an appropriate definition of the system boundaries. Most applicable was a level of system boundaries that corresponds to the fence of a production plant—the inside battery limits (ISBL).

Raw materials and energy inputs are introduced in the ISBL; products and waste streams leave the system delimited by this fence. It often includes the process plant and the corresponding utility system. Setting the system boundaries accordingly ensures that the apparatuses in scope are under direct control of the goods producer—the industrial party applying the TOP-REF methodology. An extension of this “producer’s view” would be a cradle-to-gate system boundary, including all the processes from the raw material extraction, transportation and transformation steps to the output gate of the production plant under consideration.

To model the units of a process in detail, commercial process simulation software can be used. A validation of the model developed with real plant data is necessary to ensure the correct description of the production plant by the model.

Surrogate model generation

The detailed process models were used to generate a surrogate model of the overall plant under consideration. The aim was to reduce the simulation time while increasing the robustness of the submodels, and to possibly integrate the submodels of a process plant and a utilities system.

FIG. 2. Illustration of the approach for surrogate model generation.

The process of generating surrogate models consists of linking inputs and outputs of the detailed models via simpler mathematical expressions (FIG. 2). Within the TOP-REF methodology, the inputs are the process parameters (PPs), and the outputs are the KRIs and the key performance attributes (KPAs). The KPAs are attributes that are specific for each case study and that set the boundaries for the variation of the CPPs to ensure a proper plant operation in terms of costs, production rates, and product quality and safety.

The required input and output values needed were to be provided by detailed process models of the process plant and the utilities system. Therefore, the design space of PPs was explored by a Latin hypercube sampling. The generated set of these inputs was simulated using the detailed process models to determine the corresponding outputs for each set of inputs. Based on these sets of inputs and outputs, a surrogate model generation by kriging was applied.^2,3 For each case study, a minimum of 100 converged simulation scenarios was used as input for the surrogate model generation. Non-converging scenarios were discarded and not further considered. The quality of the resulting surrogate model was improved by adding more simulation points and process knowledge.

Next to improving the simulation time and increasing the robustness required for the following global sensitivity analysis and the optimization, the surrogate model generation was also used to couple the subsystem models (e.g., the process plant and the utilities system). This coupling served the master/slave principle, as the utilities system needs to provide the energy requirements (such as steam) of the process plant. To couple both models, the interaction variables had to be identified. Based on the described master/slave principle, these interaction variables were the steam or electricity requirements of the process plant, which need to be delivered by the utilities system. By incorporating the range of the resulting interaction variable values of the process model as input for the surrogate model generation of the utilities system, a coupling of both submodels was possible. Thus, surrogate model generation was not used just for decreasing the required computational time and increasing the robustness of the submodels, but it was also instrumental as the basis for model coupling.

Global sensitivity analysis

To identify the PPs with the major influence on the KRIs and to reduce the number of operating parameters that were considered in the optimization to the most influential operating parameters (the CPPs), a global sensitivity analysis was used. Therefore, many model evaluations were needed. The generated surrogate model drastically reduced the execution times of the model evaluation runs and enabled the application of a global sensitivity analysis to identify the CPPs.

The field of sensitivity analysis aims at the characterization and quantification of the variation of a model output related to changes in the input parameters.⁴ Within the TOP-REF project, the inputs were the PPs of the considered case studies. The KRIs were the corresponding model outputs whereon the influence of the input parameters would be quantified.

Local or one-factor-at-a-time analyses rely on assumptions of model linearity and additivity. In contrast, global sensitivity analysis methods overcome these limitations by considering the whole range of input parameter variations; sensitivity estimates of individual inputs are obtained while all inputs vary simultaneously.^5,6 The interactions of different input parameters should be considered.

FIG. 3. Approach to optimize the resource efficiency within the TOP-REF project.

Variance-based methods are attractive, since they can cope with the non-linearity and non-additive characteristics of a model and can handle the interactions of the input parameters that influence the model output. A major drawback is their high computational cost for the calculation of the sensitivity measures.^4,6 Due to the aforementioned advantages of the Sobol method, a variance-based global sensitivity analysis, which was implemented in Excel VBA, was used within the TOP-REF methodology. A detailed description of the global sensitivity analysis used can be found in literature.⁷ It is based on a quasi-random Sobol sampling using an Owen-like scrambling to enable statistical error estimations.

The sampling of matrix (A + B) was performed in one and then separated to get matrix A and matrix B. Thereby, the Owen-like scrambling was also applied to matrix (A + B). The resulting first-order sensitivity index represents the individual effect of the operating parameter considered by measuring the main effect on the model output; therefore, it is the most meaningful index. The total order index is defined as the sum of all sensitivity indices; hence, it contains the individual effect connected with the interactions of the other parameters.⁸ The interactions can be calculated by subtracting the individual index from the total sensitivity index.⁹

Optimization

In terms of resource efficiency, the aim of the optimization is to find better operating points for the production plant under investigation. Therefore, the CPPs identified were used as optimization variables, the KPAs served as constraints and the KRIs served as objective functions. A schematic illustration of the approach to optimize the resource efficiency within the TOP-REF project is shown in FIG. 3. The search space of the optimization was restricted by the KPAs to ensure an optimization in a realistic and applicable range. These KPAs are specific for each case study and set the boundaries for the variation of the CPPs to ensure a proper plant operation in terms of costs, production rates, and product quality and safety, as can be seen schematically in FIG. 3 for one KRI and one CPP. Since the KPAs are outputs as the production costs or product purity, they cannot be directly related to the CPPs.

The resulting optimum might not be the global optimum, but it is an optimum within the optimizational constraints set by the KPAs. Two optimization approaches were performed. On one hand, there is a multi-objective optimization aiming to minimize all conventional KRIs (UMC, DPEC, GWU and NWU) at the same time. On the other hand, a single-objective optimization for minimizing the novel and aggregated REI was applied. Thereby, the multi-objective optimization of the more conventional indicators serves as a benchmark. This allowed a comparison with the single-objective optimization case of the aggregated REI.

The techniques used for the optimization were based on evolutionary computation and, more specifically, genetic algorithms (GAs). A simple GA was selected for the single-objective optimization,⁸ and a non-dominated sorting genetic algorithm II (NSGA-II)¹⁰ was selected for the multi-objective optimization. Within the project, a general and scalable optimization framework was developed to provide a tool for the definition of optimization problems, including the optimization variables as the CPPs, the constraints as the KPAs, and the KRIs as the objectives.

Part 2 of this article will be featured in the February issue. HP

References

1. European Commission, Impact Assessment Guidelines, https://ec.europa.eu/agriculture/sites/agriculture/files/sfs/documents/documents/sec2005-791_en.pdf
2. Rasmussen, C. E. and C. K. I. Williams, Gaussian Processes for Machine Learning, MIT Press, Cambridge, Massachusetts, 2006.
3. Pedregosa, F., G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot and É. Duchesnay, “Scikit-learn machine learning in Python,” Journal of Machine Learning Research, 2011.
4. Saltelli, A., M. Ratto, T. Andres, F. Campolongo, J. Cariboni, D. Gatelli, M. Saisana and S. Tarantola, Global Sensitivity Analysis. The Primer, John Wiley & Sons, Chichester, UK, 2008.
5. Ascough II, J. C., T. R. Green, L. Ma and L. R. Ahjua (Eds.), “Key criteria and selection of sensitivity analysis methods applied to natural resource models,” Melbourne, Victoria, Australia, 2005.
6. Saltelli, A., S. Tarantola, F. Campolongo and M. Ratto, Sensitivity Analysis in Practice: A Guide to Assessing Scientific Models, John Wiley & Sons, Chichester, UK, 2004.
7. Radatz, H., J. M. Elischewski, M. Heitmann, G. Schembecker and C. Bramsiepe, “Design of equipment modules for flexibility,” Chemical Engineering Science, 2017.
8. Boussaïd, I., J. Lepagnot and P. Siarry, A Survey on Optimization Metaheuristics, Elsevier, 2013.
9. Saltelli, A., K. Chan and E. M. Scott, Mathematical and Statistical Methods for Sensitivity Analysis, John Wiley & Sons, Chichester, UK, 2000.
10. Deb, K., A. Pratap, S. Agarwal and T. Meyarivan, “A fast and elitist multi-objective genetic algorithm: NSGA-II,” IEEE Transactions on Evolutionary Computation 6, 2002.

The Authors

Related Articles

From the Archive

Comments

Comments

{{ error }}

Add Comment

{{ comment.comment.Name }} Editor • {{ comment.timeAgo }}

{{ comment.comment.Text }}