To overcome the real-time computation and scalability issues caused by the high-dimensional decision space, nonlinearity in the generation optimization of large-scale wind-farm clusters, this paper proposes a Two-tier hierarchical framework that exploits inter-farm wake steering. Specifically, at the lower tier, each wind farm follows the power reference issued by the upper level while at the same time trying to maximise the wake-deficit velocity that reaches the inflow of downstream farms by adjusting the yaw angle of each turbine. Deep neural networks are trained to learn farm-level optimization and calculate the inter-farm wake, and act as interfarm wake models for upper tier optimization problem. At the upper tier, the controller manages the wake interactions among neighboring farms through regulating the power set-points of each farm under the given ambient wind conditions, to maximise the total power output of the whole cluster. Numerical experiments performed on a Chinese wind farm cluster confirm that the proposed framework achieves the desired optimization performance while significantly lowering computational cost and enhancing scalability compared with conventional centralized optimization.
Accurate prediction of CO₂ loading in aqueous amine solvents is essential for the design and optimization of carbon capture processes. In this study, five machine-learning models, including Decision Tree (DT), Support Vector Machine (SVM), Gaussian Process Regression (GPR), Multilayer Perceptron (MLP), and Radial Basis Function (RBF) neural networks, were developed to predict CO₂ loading in aqueous diisopropanolamine (DIPA)-based amine blends. A comprehensive dataset comprising 729 experimental data points was collected from the literature, covering temperatures of 297.8–393.1 K, CO₂ partial pressures of 0.12–5888 kPa, and amine concentrations of 10–45 wt.%. The predictive performance of the developed models was evaluated using the coefficient of determination (R²), average absolute relative deviation (AARD), and root mean square error (RMSE). Among the investigated models, the RBF network exhibited the highest prediction accuracy, achieving an R² of 0.9706, an AARD of 5.74%, and an RMSE of 0.0501. Relevancy factor analysis identified CO₂ partial pressure as the most influential parameter, while temperature showed a negative effect on CO₂ loading. The proposed machine-learning framework provides an accurate and computationally efficient tool for predicting CO₂ loading in aqueous DIPA-based amine blends, offering valuable support for solvent screening and carbon capture process optimization.
Large language models (LLMs) have attracted growing interest for process automation due to their capabilities in reasoning, code generation, and natural-language interaction. However, whether linguistic intelligence can be safely translated into physical control authority remains unclear. This study investigates this question through a progressive experimental framework, including direct LLM control, numeric-constrained LLM control, hybrid LLM-PID (Proportional-Integral-Derivative) supervisory control in a single-input single-output (SISO) water-tank system, and extension to a multivariable Continuous Stirred-Tank Reactor (CSTR). Rather than evaluating controller performance alone, the experiments were designed to identify the failure boundaries that emerge as increasing levels of control authority are assigned to an LLM. The results reveal five risks: symbolic-to-physical, physics-grounding, control stability, temporal compatibility, and multivariable reasoning. These risks are further distinguished as LLM-native risks and LLM-amplified risks, further benchmarked with the general data-driven control risks. The findings suggest that current LLMs are better suited for supervisory and decision-support roles than direct process control, highlighting the need for physics-grounded and safety-engineered AI control architectures.
Ionic liquids (ILs) are a promising class of sustainable solvents due to their negligible vapor pressure, high thermal and chemical stability, and exceptional tunability. These unique physicochemical properties have attracted considerable attention in industrial and academic fields, particularly in catalysis, separation processes, electrochemical systems, and green chemistry. Among the key properties of ILs, the refractive index (RI) is an important optical parameter that reflects molecular interactions, purity, and solvation behavior. However, experimental determination of RI is often expensive and time-consuming, especially for large datasets involving numerous IL structures under varying conditions. In this work, three machine learning (ML) techniques, including Extra Trees (ET), Elman Neural Network (ENN), and Cascaded Forward Neural Network (CFNN), were employed to predict the RI of ILs using a comprehensive dataset containing >6000 experimental data points covering 483 IL structures. The CFNN model demonstrated the best performance, achieving a root mean square error (RMSE) of 0.0020 and a coefficient of determination (R2) of 0.9977. Unlike some recent studies that rely on different sizes of datasets and focus primarily on predictive accuracy, the present work ensures a more comprehensive framework that integrates multiple complementary ML architectures with thorough interpretability and physical consistency analyses. In particular, this study goes beyond accuracy improvements by incorporating SHAP-based explainable artificial intelligence (XAI), detailed trend analysis, and leverage-based applicability domain evaluation to ensure both transparency and reliability of predictions. The proposed CFNN framework provides a reliable and transparent tool for accelerating IL screening, design, and application in green solvent engineering.
Jet flooding is a faulty condition that can occur in tray columns designed for gas-liquid separation processes. It represents a critical hydraulic operational limit for the trays, as it affects not only the separation efficiency, but can also lead to severe instability and undesired process shutdown if not promptly diagnosed. This work proposes a Depthwise Separable Convolutional Neural Network (DeptSepCNN)-based model for jet flooding detection in a laboratory-scale sieve tray glass column. We trained several DeptSepCNN architectures to solve the binary classification task in supervised learning, following an iterative procedure by exploiting a properly designed hyperparameter space and an incrementally larger number of samples. The final classifier selection was primarily guided by its performance in terms of flooding recall and F1-score, as these metrics comprehensively assess the model’s ability to correctly detect fault conditions while accounting for its robustness against false alarms. The evaluation on multiple independent test sets indicated that the final DeptSepCNN classifier achieves very high performance, especially compared to traditional CNN models, such as Vanilla CNN and MobileNet. Additionally, a Graphical User Interface (GUI) was integrated into the obtained jet flooding detector framework as an offline supervisory tool to enable near real-time classification in typical working conditions.
Textile dyeing is resource-intensive, consuming large amounts of water, chemicals, and energy due to repeated trial dyeing for color matching. Data-driven recipe prediction offers a promising route to reduce these dyeing trials, but most studies have focused on forward prediction within interpolation ranges and have rarely addressed inverse design, particularly in extrapolation regions. Here, we propose an extrapolative machine learning framework for sustainable dyeing recipe design using Gaussian process regression (GPR). While forward prediction estimates color coordinates from a given dye recipe, the inverse problem addressed here predicts dye concentrations directly from target CIELAB color coordinates. An improved GPR framework integrating inverse and forward models with target-fitting optimization and physically constrained recipe search was developed to enhance extrapolation stability. The framework was based on 250 experimental dyeing samples of microfiber fabrics dyed with three disperse dyes. To evaluate extrapolation performance, six target colors outside the training data distribution were designed and validated through actual dyeing experiments. The improved framework maintained interpolation accuracy comparable to that of the baseline inverse GPR model, which does not include target-fitting optimization. For the six extrapolation targets, it reduced the average experimental ΔEab* from 2.49 to 1.21, with all target-specific values not exceeding 2.05. These results demonstrate that extrapolative machine learning combined with experimental validation enables reliable dyeing recipe design beyond existing datasets. By reducing trial dyeing and improving prediction reliability, the proposed framework offers a practical pathway toward resource-efficient and sustainable textile dyeing.
Flushing operations in multiproduct lubricant oil pipelines are a critical determinant of product purity, operational efficiency, and economic performance, yet traditional approaches often rely on operator experience and trial-and-error, leading to excessive waste and frequent cross-contamination. This research presents an integrated framework combining data-driven analysis, experimental validation, and process optimization to address these challenges and optimize packaging operations in petroleum industries. Industrial data capturing both successful and failed flushes were analyzed using a suite of candidate machine learning (ML) classification models including ensemble methods, kernel-based models, and deep learning architectures to accurately classify flushing outcomes and identify the key factors driving successful flush. To mitigate the significant class imbalance between successful and unsuccessful flushes, Synthetic Minority Oversampling Technique (SMOTE) was employed, ensuring robust model performance. SHAP analysis revealed that flushing success is governed by complex, nonlinear interactions among lubricant type, viscosity contrasts, pipeline flow dynamics, flushing volume, and ambient temperature, emphasizing that no single factor dominates the process. Insights from this analysis informed the design of a bench-scale experimental rig that replicated industrial pipeline geometry and hydrodynamics, enabling systematic evaluation of improved drainage procedures, air-assisted flushing, and flow behavior. These results guided the development of a novel standardized and optimized flushing protocol, which, when implemented in a production line, reduced the failure rate from 11% to 3.6%, minimized flush volumes, and delivered significant cost savings. This study highlights the synergy between data-driven insights, experimental validation, and process optimization in achieving high-impact outcomes that are resilient, scalable, and sustainable. The framework not only enhances operational reliability in petroleum packaging operations but also provides a generalizable methodology applicable across process industries seeking to improve contamination control, efficiency, and economic performance.
The rapid digitalisation of chemical engineering practice has created an urgent need for educational frameworks that ensure graduates are equipped with the digital skills required by industry. This paper presents the development of a new set of Competencies for Digital Skills (CDSs) for chemical engineering, designed to integrate seamlessly into undergraduate, master’s, and continuing professional development programmes. Drawing from a review of accreditation learning outcomes, professional standards, international accreditation guidance, sector reports, literature, and programme specifications, over 40 digital-related Intended Learning Outcomes (ILOs) were identified and thematically grouped into four key competency areas: Data-Driven Engineering and Digital Solutions; Systems Thinking and Process Optimisation; Professional Ethics, Risk, and Compliance; and Communication and Engineering Project Delivery. These were synthesised into a structured set of sixteen practical CDSs, each supported by example progressive proficiency descriptors aligned with European Qualifications Framework (EQF) levels 6–7 and beyond. Rather than treating emerging technologies such as machine learning and AI as isolated competencies, the framework embeds them as tools that enhance capability across the digital skillset. This work provides a practical and forward-looking foundation for embedding digital capability across chemical engineering education.
Activated sludge models are essential for wastewater-treatment design and operation, but their detailed simulations can be too computationally demanding for repeated analysis. Machine-learning surrogates are faster, yet accurate predictions may still violate mass conservation or produce negative component concentrations. This study introduces Invariant-Constrained Second-Order Regression (ICSOR), an interpretable surrogate that combines a transparent, invariant-aware regression head with a constrained deployment procedure. The regression head predicts the effluent component state from influent conditions and operating inputs. At deployment, each prediction is checked, projected onto the mass-conservation constraints, and, when necessary, adjusted by a linear program to enforce both conservation and non-negativity. ICSOR was evaluated on 10,000 steady-state samples from an ASM2d-TSN continuous stirred tank reactor and compared with XGBoost, LightGBM, CatBoost, AdaBoost, Random Forest, support vector regression, k-nearest neighbors, partial least squares, and a multilayer perceptron. The deployed ICSOR predictions had zero conservation and non-negativity violations: the affine projection resolved 21.8% of test cases, and the linear program resolved the remaining 78.2%. For context, none of the raw regression-head predictions satisfied conservation, although 21.7% were non-negative, showing that the hard guarantee came from constrained deployment rather than training alone. This physical reliability involved a moderate accuracy tradeoff: aggregate test RMSE was 5.98, compared with 4.38 for the multilayer perceptron and 5.30 for LightGBM. Across the dataset-size study, however, ICSOR achieved the lowest normalized area under the RMSE learning curve and the smallest train–validation RMSE gap among the comparatively accurate models. ICSOR therefore offers a useful balance of interpretability, data efficiency, and guaranteed physical admissibility for steady-state activated sludge surrogate modeling.
Polyhydroxyalkanoates (PHAs) are biodegradable polyesters produced by microorganisms and represent one of the most promising alternatives to petroleum-derived plastics. However, its commercialization is constrained by high production costs depending on the choices of microbial strain and carbon substrate, yet experimental investigations of new strain–substrate combinations remain resource-intensive. This study presents a substrate-strain-agnostic, robust, hybrid framework, combining a dynamic, mass-balance and kinetic-based mechanistic model (DPM) and machine learning (ML), to predict fermentation performance in terms of cell biomass and PHA concentrations from any strain–substrate combination. The DPM-generated data and experimental time-course profiles spanning sugar-based substrates, oils, and waste carbon resources, including lignocellulose and waste cooking oils, have been used to train/test ML models. The ML models (with 80:20 train:test split) thus built capture inherent dynamic interactions among carbon source depletion, nitrogen limitation, biomass proliferation, and intracellular PHA-copolymer accumulation. Artificial neural network (ANN) and random forest (RF) surrogates predict cell biomass and PHA concentrations from fermentation-state features: time, starting and current substrate and nitrogen concentrations. Both surrogates achieved R² > 0.99 on the overall test set. However, the ANN outperformed the RF on the unseen experimental profiles (R² > 0.95 vs > 0.80). Thus, the ANN model, available as open-source software: https://jhumasadhukhan.github.io/ANN-model-to-PHA-prediction/, is recommended for strain–substrate screening, batch feed-and-nitrogen-limit design, and soft-sensing in digital-twin-enabled fermentation.
The Tennessee Eastman Process is a widely used benchmark for process control and fault detection and diagnosis, yet earlier implementations in FORTRAN and MATLAB C-mex restricted access to internal variables, limiting flexibility for advanced studies. The recently developed complete and open Simulink model of the TEP (COSTEP), removes these barriers by providing full access to all process parameters and expanding the number of available measurements to 145. This open architecture makes COSTEP particularly valuable for customised dataset generation, process monitoring, and machine learning applications aligned with the goals of Industry 4.0 and Big Data.This paper demonstrates the advantages of COSTEP as an open, accessible, and extensible simulation framework for the Tennessee Eastman Process (TEP). Rather than focusing primarily on the development of a new fault detection method, the study illustrates how the additional transparency and flexibility provided by COSTEP enable analyses that are difficult to perform using previous implementations. This is demonstrated through two case studies. First, a review of all the Tennessee Eastman Process fault conditions shows how direct access to internal process variables, manipulated variables, and actuator signals allows nominal benchmark faults to be examined in greater detail. This enables a clearer distinction between faults that produce meaningful process deviations and cases that may represent weakly expressed faults or normal closed-loop operating variations. Second, exergy analysis is implemented as an example of an advanced user-defined analysis made feasible by COSTEP’s comprehensive access to process information. Together, these case studies demonstrate that COSTEP is not only a simulation model, but a flexible benchmark platform for process understanding, data generation, machine learning, optimisation, and fault detection and diagnosis research.
This paper presents the design and evaluation of a Graph-RAG based chatbot developed to assist process safety engineers in assessing the safety position of industrial plants. Currently, process safety engineers must rely on manual collection and interpretation of data, which is both time-consuming and prone to human error. The introduction of a chatbot-based solution has the potential to streamline this process by providing rapid, consistent, and data-driven insights, benefiting both individual engineers and organisational decision-making as a whole. The chatbot integrates data from multiple sources-including Barrier Safety Panels (BSPs), work order databases, and also Failure Modes, Effects, and Criticality Analyses (FMECA)-into a unified interface. By combining structured knowledge stored in a graph database model with real-time status data in SQL databases, the system provides engineers with contextualised, accurate responses to complex operational questions. The evaluation demonstrates that while the chatbot successfully integrates and interprets heterogeneous datasets, challenges remain in numerical and date processing, prompt adherence, and the mitigation of hallucinations. Across a structured set of test queries, the system consistently returned relevant data, although inconsistencies in formatting and occasional hallucinated responses were observed. Overall, this proof of concept shows that Graph-RAG can significantly enhance situational awareness and decision-making in safety-critical industries, but further refinement is required to ensure robustness, consistency, and trustworthiness in realworld deployments.
This study presents simulations of integrated upstream and downstream monoclonal antibody (mAb) production processes using embedded high-resolution modules for the cultivation and capture of chromatography units. To reduce computational costs, a surrogate model was developed and validated, resulting in a 60,000-fold reduction in computational time while maintaining high fidelity to the physical models. Sensitivity analyses employing the quasi-Monte Carlo method with over 4096 simulations were conducted to assess the influence of cell line and process parameters on annual productivity and production costs. The results identified critical design parameters and emphasized the effect of scaling up on operations. Upstream operations dominated total production costs (similar to 70%), with labor decreasing from 77% at 1000 L to 36.2% at 10,000 L, and culture media accounting for 90% of the material costs. Larger reactors reduced the unit mAb cost (74.3 to 32.5 USD g(-)& sup1;), whereas shorter cultivation increased throughput but also raised costs (from 53.7 to 65.6 USD g(-)& sup1;), highlighting a trade-off between production rate and unit cost. The use of high-productivity cell lines effectively mitigates these tradeoffs. Consequently, this study provides a valuable computational tool to support decision-making in process design, optimization, and control in biopharmaceutical manufacturing.
This article presents a data-driven strategy for applying a Super-Twisting-based Extremum Seeking Control (ST-ESC) algorithm, whose goal is to maximize hydrogen production in a dark fermentation process. The methodology uses experimental design and machine learning based regression techniques to model the steady state response surface of the dark fermentation process, thereby describing the objective function of the optimization problem to be solved. The manuscript details each stage of the strategy, from experimental design and data collection to the machine learning-based modeling phase and extremum-seeking control design. It is important to note that the experimental data used in this study were not obtained from physical experiments, but rather generated through simulations of a validated mathematical model that accurately represents the real biotechnological process. The numerical results, as measured by the root mean square error (RMSE), demonstrate the effectiveness of the proposed data-driven modeling strategy and its potential for control applications. Two approaches were compared for response surface generation: (i) a non-cumulative approach, which iteratively adds five new points without retaining previous ones, and (ii) a cumulative approach, which preserves all previously added points. The non-cumulative strategy achieved higher accuracy (RMSE = 0.0001821) compared to the cumulative one (RMSE = 0.0008222), although it may involve higher computational and experimental costs. Nevertheless, both approaches effectively tune the model to reproduce the real system's behavior, which underpins the ST-ESC design. By combining these advanced techniques, the research presents a data-driven optimization framework to maximize the hydrogen production rate online.
The Data Exchange in the Process Industry (DEXPI) standardization group has recently released a new specification addressing early process design information, particularly information found in block flow diagrams (BFDs) and process flow diagrams (PFDs). The specification provides an information model and exchange format for such design information, but it does not yet define a graphical notation or modeling approach for creating and exchanging compliant BFDs or PFDs. Our work investigates the structural and semantic correspondence between the DEXPI Process information model and the Business Process Model and Notation 2.0 (BPMN 2.0) standard and proposes a BPMN 2.0-based representation for DEXPI’s BFDs and PFDs. The approach is validated using the Tennessee Eastman process as a case study. The results show that BPMN 2.0 can represent and exchange DEXPI Process-relevant design information in a semantic and standardized way, that the resulting models can be transformed into schema-valid DEXPI 2.0 XML, and that it provides a practical and tool-supported basis for graphical modeling of BFDs and PFDs. To support adoption, we provide an open-source reference implementation, bpmn2dexpi, that realizes the proposed representation, transforms BPMN 2.0 models into DEXPI 2.0-compliant XML, and validates the output against the DEXPI 2.0 XML Schema and information model.
In many cyclic process operations a phase switch is triggered when a measured signal crosses a plant-defined threshold, and anticipating that crossing rather than reacting to it can yield significant energy and product-quality gains. The decision therefore reduces to forecasting the time-to-threshold (TTT), the time remaining until a cyclic signal crosses the threshold, from a partially observed cycle. We present a Multi-Feature k-Nearest Neighbours (MF-kNN) framework with Trajectory-Embedded Decay (TED) for real-time TTT forecasting in such operations. The framework performs direct, one-shot TTT prediction by matching the partially observed current trajectory to historical cycle segments across multiple process variables and reusing their future continuations, thereby avoiding recursive rollouts and model retraining. Its central component, TED, adapts the forecast online by reweighting and pruning neighbours according to exponentially decayed rolling prediction errors, giving segment-aware adaptation without per-segment retuning. To ensure methodological rigour and commercial deployability, the framework further incorporates boundary-safe cycle segmentation, cycle-consistent windowed neighbour search, moving-window embeddings, and clustering-guided mode-aware memory restriction. Applied to 52 industrial regeneration cycles under a cycle-blocked evaluation protocol, MF-kNN outperformed univariate k-NN and a panel of statistical, neural, and lightweight machine-learning baselines (ARMA, Theta, LSTM, ridge-regularised autoregression, and gradient-boosted trees). Relative to univariate k-NN, the multivariate formulation reduced the mean TTT error by approximately 20 % and the trajectory error by approximately 12 %, while producing markedly more stable forecasts across cycles; relative to the strongest trained baseline (LSTM) it reduced TTT error by approximately 79 % (nearly fivefold). Windowed search delivered a 4.9× per-forecast speedup with no measurable loss of accuracy, and TED further reduced the mean TTT error by approximately 48 % relative to baseline MF-kNN. These results show that the proposed framework provides a transparent, efficient, and practically deployable solution for long-horizon cycle-time forecasting in cyclic adsorption processes.
Mechanistic models of catalytic reactors are robust but often fail to generalize across feedstocks and operating conditions due to incomplete kinetic representations and parameter uncertainty, while purely data-driven models lack physical consistency and require large datasets that are costly to obtain in process engineering. To address these challenges, this work proposes a physics-informed hybrid framework that integrates a kinetic ordinary differential equation (ODE) into a neural network through automatic differentiation, differential residual loss minimization, and physical boundary constraints, taking both operating conditions and feedstock descriptors as inputs. The approach is applied to industrial hydrotreating (HDT), focusing on the prediction of nitrogen slip concentration in vacuum gas oil (VGO) hydrodenitrogenation (HDN).Two methodological developments are proposed: a Delaunay-triangulation-based strategy to select the input locations at which the physical residual and boundary terms are enforced during hybrid training, and a feedstock oriented evaluation across 30 randomized splits in which entire feedstocks are withheld from training to assess extrapolation. The hybrid model consistently outperforms both the mechanistic and purely data-driven baselines, achieving lower RMSE, MAE, and temperature deviations, while trend analyses confirm stronger preservation of the underlying physical dependencies, even under extrapolation to unseen feedstocks. A systematic comparison using a simplified kinetic formulation further shows that the hybrid architecture remains effective when constrained by incomplete mechanistic knowledge, outperforming data-driven models and approaching the accuracy of the complete kinetic expression.These results highlight the robustness, flexibility, and practical relevance of physics-informed hybrid modeling for reactor systems, offering a promising pathway toward reliable and quickly developed digital twins in catalytic process engineering.
This work aims to compare data-driven, first-principles and hybrid modelling approaches for the prediction of a zeolite product particle size distribution (PSD) in a jet mill process. A small dataset is collected from jet milling experiments with different process conditions and feed PSDs. A population balance model (PBM) serves as the first-principles model, while the machine learning model is an artificial neural network (ANN) using Monte Carlo dropout, predicting the final PSD directly from the process conditions and feed PSD. A serial hybrid model (SHM) is implemented using an ANN to predict two PBM parameters, and as well as a corrective hybrid model (CHM) using an ANN to predict the PBM error. All of the models are trained and evaluated in a leave-one-out cross-validation (LOOCV) procedure, and hyperparameters are optimised using Optuna.The SHM is found to be better than the PBM alone but worse than the other purely data-driven models, suggesting that the chosen PBM structure is too simple to capture the complexities of the jet milling process for the zeolite material, and thus constrain and limit the data-driven component of the serial model. The CHM is the best performing model, achieving a 34.5% reduction in cross-validation mean absolute error (MAECV) over the purely data-driven model. The CHM performance is explored on the two held-out test experiments, showing the effective corrective action on the PBM predictions with uncertainty estimates from Monte Carlo dropout. Explainable AI techniques (Shapley values) are implemented to understand the jet mill parameters which have significant influence on the magnitude and direction of the predicted corrective action.The work demonstrates the advantages of hybrid modelling, and highlights the importance of selecting an appropriate hybrid model structure with consideration of the underlying first-principles model’s capabilities and limitations. It demonstrates the application of explainable AI techniques with corrective hybrid modelling for development of the underlying first-principles model.
Increasing greenhouse gas emissions have intensified efforts to replace fossil fuels with renewable alternatives. Biodiesel offers reduced emissions compared to fossil diesel. However, the biodiesel production feasibility, design, and optimization show a strong influence of the liquid-liquid equilibrium (LLE). This work compared three machine learning models, an artificial neural network (ANN), a Random Forest, and an XGBoost, against a correlated UNIFAC Dortmund thermodynamic model for predicting LLE in biodiesel systems. For the model’s correlation, 1,119 experimental tie lines compiled from 54 peer-reviewed papers were employed. All machine learning models significantly outperformed the baseline, with XGBoost achieving the highest accuracy (RMSE of 0.0093), representing an 84% improvement over UNIFAC Dortmund predictions. Additionally, a sensitivity analysis was implemented to understand the temperature and pseudomolecule molecular weight effects on the LLE. The XGBoost model revealed strong temperature and molecular weight influence across all scenarios. Low values for these variables promote the component of interest recovery. While machine learning models showed superior predictive accuracy, their black-box nature precludes a mechanistic understanding of molecular interactions and thermodynamic property estimation. Traditional phenomenological models, despite lower accuracy, provide the best interpretability for the LLE.
To address the non-linearity and large time-delay characteristics of selective catalytic reduction (SCR) denitrification system, this study proposes an enhanced twin delayed deep deterministic policy gradient (ETD3) algorithm. Key improvements to the TD3 algorithm include: (1) integrating a multi-step temporal difference (MSTD) approach to compute target Q-values using historical state sequences, thereby improving handling of delayed rewards and system dynamics under time delays; (2) designing an adaptive delayed update mechanism for the Actor network to dynamically adjust update intervals, reducing control lag during load fluctuations. Additionally, a SCR denitrification control strategy based on ETD3 algorithm is proposed, while theoretical analysis, based on the Hamilton-Jacobi-Bellman (HJB) equation, is conducted to verify the stability of the ETD3 proposed controller. For varying operating conditions, the control process is divided into three load-based stages. The ETD3 controller is trained across all three stages using corresponding SCR denitrification models, and a fuzzy gain-scheduling controller linearly combines these models to better represent denitrification system nonlinearity. Simulation and industrial application results confirm that the ETD3-based strategy offers superior disturbance rejection and robustness under varying operating conditions.
