Conference Agenda

In catalytic processes such as Steam Methane Reforming (SMR), multiple parallel and competing reactions occur, influencing product yields and reactor efficiency. The objective of this work is to develop a methodology based on reinforcement learning (RL) to accurately map the most probable reaction pathways by utilizing experimental data, such as partial pressures of methane (CH₄), hydrogen (H₂), carbon monoxide (CO), and carbon dioxide (CO₂) measured over time and temperature. By leveraging this data, the RL model dynamically selects the reaction pathways that best reflect the underlying reaction kinetics and mechanisms, distinguishing itself from conventional deterministic methods used in the literature.

Unlike traditional reaction modeling approaches, which often rely on predefined mechanisms, this methodology allows the RL agent to explore the reaction space autonomously. It considers a wide array of reactions, including Steam Methane Reforming, Water-Gas Shift (WGS), Dry Methane Reforming (DRM), methane decomposition, the Boudouard reaction, methanation, reverse WGS, and CO hydrogenation. The RL agent is trained using a Q-learning framework with an ε-greedy exploration-exploitation policy, which balances the search for new reaction pathways (exploration) and the optimization of the best-known reactions (exploitation). The algorithm optimizes the selection of these pathways by iteratively improving the match between predicted gas compositions and experimental data, learning which reactions dominate under specific conditions, such as varying temperatures and residence times.

The model is designed to incorporate data-driven adaptability into the pathway synthesis, enabling it to select the optimal reaction scheme that best reflects the behavior of the reactive system under varying operational conditions, such as changes in temperature, pressure, and residence time. This real-time adaptability is crucial for accurately capturing the dynamic nature of catalytic processes, which traditional deterministic models often struggle to account for. Furthermore, the RL model employs a reward function that penalizes the selection of pathways that are either infeasible or deviate from established thermodynamic principles, ensuring that the reaction networks remain physically consistent while accurately representing known reaction kinetics.

Initial results show that this RL-based pathway selection significantly reduces prediction errors and enhances the identification of dominant reaction mechanisms, particularly in complex systems where multiple parallel reactions compete. The model’s ability to adjust dynamically to experimental data demonstrates its superiority over classical methods, providing a flexible and robust tool for optimizing catalytic processes like SMR.

This study demonstrates how the integration of Reinforcement Learning with Reaction Engineering can enhance the understanding and prediction of reaction pathways, offering a scalable solution for both research and industrial applications in reaction mechanism optimization.

This work subsists on the simulation and optimisation modelling of an integrated Heat Recovery System (HRS) for a petrochemical plant. The conceptualised system in this work includes four combustion-based processes (CBP) (two using natural gas as fuel and other two using liquified petroleum gas) and a condensing steam turbine combined heat and power generation (ST-CHP) system. The conceptualised system furtherly incorporates technologies to recover heat from the exhaust gases of the combustion-based processes, resulting in fuel savings (through preheating of combustion air at the inlet of the four CBP’s and of the water at the inlet of the ST-CHP’s boiler), as well as electricity generation via an organic Rankine cycle (ORC).

An existing methodology (Castro Oliveira, 2023) approaching the highest level of waste heat stream recirculation between a determinate number of production processes is expanded to include energy supply systems (a category in which CHP is included) as well. In this sense, this work introduces an innovative perspective in which the production processes in a plant (such as CBP’s) and the energy-using units within energy supply systems may be analysed in a similar manner due to the possibility to incorporate of the same type of improvement technologies (in this case, heat exchangers) within the overall systems (and thus to attend to the most fundamental aim of the energy efficiency improvement of the plant). The objectives of this work are framed on the aims of the EU Strategy on Energy System Integration (European Commision, 2020), namely its first pillar (energy efficiency and circular economy nexus).

The simulation model for the proposed system was developed using the Modelica language, using the capabilities of the ThermWatt Modelica library (Castro Oliveira, 2023), which has been specifically designed for the development of models of energy recovery and water systems. An optimisation model using the non-linear programming (NLP) methodology was then developed based on this second scenario (using the same set of variables and governing equations of the corresponding simulation model). The Python language-based GEKKO optimisation package was used to build this model. The objective was defined as the minimisation of operational costs associated to the plant’s energy consumption (including both fuel and electricity consumption).

A post-processing assessment was furtherly performed subsisting on the determination of the economic and environmental viability associated to the engineering project of the conceptualised system associated to 108.9 TJ/year total fuel savings (corresponding to a relative 13.8% reduction) and 147.2 MWh/year electricity savings. In this prospect, a payback time of about 2 year and 3 months and a 5.14 kt/year reduction of equivalent carbon dioxide (CO_2,eq) emissions have been estimated, which are significantly reasonable values in comparison to respective benchmarks 2 – 3 years reasonable payback time (Tello and Weerdmeester, 2013) and a 12 – 19 kt CO_2,eq /year reduction (ABB, 2023).

References

M. Castro Oliveira, 2023, ThermWatt Home Page, https://fenix.tecnico.ulisboa.pt/homepage/ist178789/thermwatt---ferramenta-amp-servico-de-engenharia.

European Commision, 2020, Powering a climate-neutral economy: An EU Strategy for Energy System Integration.

P. Tello, R. Weerdmeester, 2013, Spire Roadmap, 106.

ABB, 2023, Energy efficiency opportunities in chemical manufacturing.

The complexity of describing multiphase systems universally has led to the development of various models tailored to specific industrial flow scenarios (Oran & Boris, 2002). Multiphase flows, commonly found in industries such as oil, gas, and chemical processing, require systems to operate within specific flow regimes. When these regimes are violated, systems can fail, making it critical to identify flow regimes accurately to maintain operational safety and efficiency (Xu et al., 2022).

Existing techniques for identifying flow regimes often rely on data from real systems, requiring expensive infrastructure and offering limited adaptability to new conditions. As an alternative, this study uses computational fluid dynamics (CFD) simulations, which allow for controlled experiments without the need for costly equipment. CFD was applied to simulate water-oil mixtures with different flow regimes, including slug, annular, and dispersed bubble patterns, providing data for flow classification. Typically, multiple sensors or variables are used for flow classification (Wang et al., 2019; El-Sebakhy, 2010), but to minimize costs, this study used only the apparent density in a cross-section of the simulated pipe, which could be measured by a single sensor in real systems.

To simplify the classification process, the numerical integration of the density was proposed as a one-dimensional variable representing flow characteristics. This simplified representation allowed for the use of machine learning techniques, and k-means was chosen as the classification method. Traditional classification of multiphase flows is often done visually, which can lead to subjective interpretations (Wu et al., 2001). To avoid this, k-means, an unsupervised learning method, was used. K-means works by minimizing an objective function based on random cluster assignments, making it ideal for problems where labeled data is not available.

Despite the randomness of k-means, the method consistently provided accurate classifications, with none of the results showing an accuracy lower than 80%. This demonstrates that k-means can reliably differentiate between flow regimes using only the integrated density variable. The decision to use k-means was based on its ability to perform well in scenarios without predefined classification criteria, making it suitable for multiphase flow studies where labeled data is scarce or unavailable.

In conclusion, applying k-means for flow regime identification using CFD data presents a cost-effective and reliable solution. By relying on only one variable—apparent density—the method significantly reduces the need for expensive instrumentation, making it practical for real-world applications. Furthermore, this study highlights the potential of CFD simulations to provide valuable data for flow classification, offering an efficient alternative to traditional experimental methods. This approach can improve safety and operational efficiency in industries that deal with complex multiphase flows.

TRANSLATE with x English ArabicHebrewPolish BulgarianHindiPortuguese CatalanHmong DawRomanian Chinese SimplifiedHungarianRussian Chinese TraditionalIndonesianSlovak CzechItalianSlovenian DanishJapaneseSpanish DutchKlingonSwedish EnglishKoreanThai EstonianLatvianTurkish FinnishLithuanianUkrainian FrenchMalayUrdu GermanMalteseVietnamese GreekNorwegianWelsh Haitian CreolePersian TRANSLATE with COPY THE URL BELOW Back EMBED THE SNIPPET BELOW IN YOUR SITE Enable collaborative features and customize widget: Bing Webmaster Portal Back

• Adicionar ao Dicionário
  • Não há listas de palavras para Inglês → Português...
  • Criar uma nova lista de palavras...
• Copiar