R&DD
R&DD
Deep reinforcement learning optimization framework for a power generation plant considering performance and environmental issues
Achieving a balance between power production and emission limits is a complex task in the electric power generation sector. A deep reinforcement learning optimization framework (DRLOF) was developed for a circulating fluidized bed (CFB) power plant. The DRLOF utilized plant data to create an environment and interacts with an advantage actor-critic (A2C) agent. The framework maximized power generation while adhering to capacity constraints and environmental emission standards, considering the cost of operations. The separate-A2CN architecture achieved an increase in electricity generation and emission reduction with lower computational cost. The DRLOF demonstrated flexibility, adaptability, and lower computational burden through different test scenarios. This approach has implications beyond the CFB power plant, extending to other chemical processes and industries. The model contributes to optimization challenges, customization needs, and cost reduction in online operations.
CO2 capture capability of 0.5 MW MEA demo plant using three different deep learning pipelines
Three deep learning pipelines (S-DLP, SBE-DLP, and L-DLP) were developed to predict the CO2 capture potential of amine-based capture processes. Real operating data from a 0.5 MW MEA demo plant were utilized for data cleaning, feature selection, and deep neural network-based prediction. After eliminating outliers, the L-DLP was selected as the best pipeline, providing accurate predictions of CO2 capture rate and concentration. Furthermore, the L-DLP was used to predict temperature variation in the absorber and stripper, crucial for energy saving. The developed pipeline demonstrated its potential for accurate and efficient prediction of CO2 capture in chemical processes.
Techno-economic analysis and optimization using DNN model for a CO2 absorption process using an MDEA/PZ solvent with a solvent looping system for steam methane reforming
The DNN model was developed for a pre-combustion CO2 absorption process utilizing a solvent looping system with a blended MDEA/PZ solvent. The novel solvent looping system operated by recycling a portion of the solvent from a flash drum, thereby reducing energy consumption and the size of the stripper. Deep neural network (DNN) models were utilized to accurately predict capital and operating costs. A multi-variable mutation approach using DNN models was employed to optimize the operating conditions and minimize the CO2 capture cost.
Actor-critic reinforcement learning using the DNN model to determine the optimal operating conditions of the hydrocracking process
An actor-critic reinforcement learning (RL) optimization strategy using a deep neural network (DNN) surrogate model was developed for optimizing the operating conditions of hydrocracking units because traditional mathematical models for optimization are costly and may not be suitable for the quick response and customization needs of hydrocracking units. The DNN surrogate model, built from a validated mathematical model, provides accurate predictions with minimal computational resources. The proposed RL framework achieved an accuracy of 97.86% to 98.5% in determining optimal operating conditions, demonstrating consistency and an average efficiency of 98% in case studies. The approach offers advantages such as quick response time, low computational burden, and customizability, making it practical for online implementation. This strategy can be extended beyond hydrocracking to other chemical industries, presenting a novel and valuable optimization solution.
Integrating mathematical modelling and machine learning for steam methane reformer in the chemical industry
The steam methane reformer (SMR) plays a crucial role in hydrogen production using natural gas, making it attractive to the chemical industry. A combined approach using mathematical modeling and artificial neural networks (ANN) was developed to enhance SMR performance. A rigorous dynamic model for the SMR was developed and validated, demonstrating its reliability in predicting temperature, pressure, mole fraction, and heat flux. The model was then used to generate performance data by varying catalyst parameters and operating conditions. The dataset was analyzed using principle components, and an ANN was trained to accurately map the relationship between operating variables and predicted outputs. The ANN achieved high accuracy with significantly reduced computational time compared to dynamic simulations. This methodology enables online operation, optimization, and system design for hydrogen production at low computational cost, making it valuable to the industry.
ANN-based optimization for low-carbon H2 production via a sorption-enhanced steam methane reforming (SESMR) process integrated with separation process
The ANN model was developed for an SESMR process integrated with a separation process for low-carbon H2 production. The development aims to simplify optimization and control, provide guidance for sorbent-catalyst development, and leverage the potential of artificial neural networks (ANN) for accurate prediction. The integrated process consists of cyclic fluidized bed (CFB), pressure swing adsorption (PSA), and other units. Sensitivity analysis and feature selection were performed to identify main objectives and variables. An ANN-based optimization approach was developed to formulate a SESMR-driven model, resulting in a 15% cost reduction for high purity H2 production and CO2 capture compared to the traditional steam methane reforming process. The findings contribute to the operation, optimization, and control of integrated SESMR processes and have implications for sorbent and catalyst development in complex processes.
ML model of commercial scale SMB for multi-component n-paraffin separation
The ML model was developed for the analysis and improvement of simulated moving bed chromatography (SMB) for the separation of a multi-component mixture of C10-14 normal paraffins from kerosene. To address the complexities of feed composition and operational variables, a mathematical dynamic model was used alongside data-driven machine learning methods, which employed real industrial data, to evaluate SMB performance. Analysis using the developed dynamic model suggested that optimizing the countercurrent ratio in different zones could enhance recovery, and the importance of 'zone flush' was highlighted for effective impurity removal. The complementary use of mathematical and data-driven models was recommended due to the limitations of available experimental data. Furthermore, an exergy analysis was conducted to assess energy flow in the overall SMB process, providing suggestions for potential improvements. The results contribute to the development of design and operation guidelines for improving SMB performance in the separation of normal paraffins.
Design guideline for CO2 to methanol conversion process supported by generic model of various bed reactors
A novel reactor design guideline for the CO2 to methanol (CTM) conversion process was introduced by developing a generic model. The analysis encompassed commercially viable Bed Reactors (BRs), and the various optimal operating conditions were determined. Different BRs had different optimal conditions, and the overall CTM process demonstrated minimum cost at specific reactor pressure and temperature values. Developed guidelines offer practical applications for CTM process design, operation, and decision-making, and the BR models can contribute to other CO2 utilization processes.
A Novel Process for Wet CO2 to Methanol using a Catalytic Fluidized Bed Reactor integrated with Separators
This ML model and optimizer were developed for an advanced process to convert wet CO2 into methanol using a catalytic fluidized bed reactor integrated with separators. With the increasing demand for large-scale green processes to neutralize carbon emissions, this model offers an attractive solution. An artificial neural network (ANN) model could optimize the process, accurately predicting performance and significantly reducing computational costs
Deep Neural Network-based Surrogate Model for Dynamic Behaviors and Performance of Solid Propellant Combustion
The development focuses on predicting the dynamic behavior and performance of solid propellant combustion using a mathematical model and a deep neural network (DNN) surrogate model. The rigorous mathematical model has limited practicality due to its computational complexity. The DNN surrogate model was developed to address this limitation, accurately and quickly predicting combustion dynamic behaviors. The surrogate model outperformed dynamic simulations in terms of computational speed, being hundreds of times faster.
Advanced cartridge design using CFD for a gas respiratory protection system
Gas masks play a vital role in safeguarding individuals from hazardous materials, necessitating the development of advanced cartridges for next-generation gas respirators. The CFD model was validated using a prototype cartridge, and the respiration resistance and breakthrough time were compared for the three designs. A virtual reality (VR) system was also employed to observe the cartridges' performance visually. The CFD model contributes to developing and optimizing gas respirator cartridges, enhancing wearability, protection, and overall design efficiency.
