Hybrid Neural Network Models Applied to a Free Radical Polymerization Process

This article presents different ways of obtaining hybrid models, which are composed of a simplified phenomenological model and one or several neural networks. As an example, we consider free radical polymerization of methyl methacrylate, achieved through a batch bulk process, in which modeling of conversion and polymerization degrees is analyzed. Kinetics of the process is described through a simplified phenomenological model that does not take into account the gel and glass effects. This last part of the process, which is more difficult to model, is rendered by means of feed-forward neural networks with one or two hidden layers. In the present paper, the hybridization procedure is made in three ways: 1) the neural network corrects the outputs of the simplified kinetic model by modeling the residuals of conversion and polymerization degrees; 2) the neural network provides accurate values of the rate constants to the simplified kinetic model; 3) the neural network models that part of the process in which gel and glass effects appear. It is demonstrated that accurate results are obtained in all three cases, and the hybrid models are easily created and manipulated, especially because they are based on neural networks with quite simple topologies.     

Control of a Batch Polymerization System Using Hybrid Neural Network ‐ First Principle Model

In this work, the utilization of neural network in hybrid with first principle models for modelling and control of a batch polymerization process was investigated. Following the steps of the methodology, hybrid neural network (HNN) forward models and HNN inverse model of the process were first developed and then the performance of the model in direct inverse control strategy and internal model control (IMC) strategy was investigated. For comparison purposes, the performance of conventional neural network and PID controller in control was compared with the proposed HNN. The results show that HNN is able to control perfectly for both set points tracking and disturbance rejection studies.     

Neural Network Based Control of the Acrylonitrile Polymerization Process

Acrylic fiber is commercially produced by free radical polymerization, initiated by a redox system. Industrial production of polyacrylonitrile is a variant of aqueous dispersion polymerization, which takes place in a homogenous phase under isothermal conditions with perfect mixing. The fact that the kinetics are a lot more complicated than those of ordinary polymerization systems makes it difficult to control the molecular weight. On the other hand, abundant data is being gathered in industrial polymerization systems, and this information makes the neural network based controllers a good candidate for managing such a difficult control problem. Multilayer neural networks have been applied successfully in the identification and control of dynamic systems. In this work, the neural network based control of continuous acrylonitrile (ACN) polymerization is studied, based on a previously developed new rigorous dynamic model for the polymerization of acrylonitrile. Two typical neural network controllers are investigated, i.e., model predictive control and NARMA‐L2 (Nonlinear Auto Regressive Moving Average) control. These controllers are representative of the variety of common ways in which multilayer networks are used in control systems. The results present a comparison of the two common neural network controllers, and indicate that the model predictive controller requires a larger computational time.     

An algorithm to determine sample sizes for optimization with artificial neural networks

This article presents an algorithm developed to determine the appropriate sample size for constructing accurate artificial neural networks as surrogate models in optimization problems. In the algorithm, two model evaluation methods—cross‐validation and/or bootstrapping—are used to estimate the performance of various networks constructed with different sample sizes. The optimization of a CO₂ capture process with aqueous amines is used as the case study to illustrate the application of the algorithm. The output of the algorithm—the network constructed using the appropriate sample size—is used in a process synthesis optimization problem to test its accuracy. The results show that the model evaluation methods are successful in identifying the general trends of the underlying model and that objective function value of the optimum solution calculated using the surrogate model is within 1% of the actual value. © 2012 American Institute of Chemical Engineers AIChE J, 59: 805–812, 2013     

Modelling and control of different types of polymerization processes using neural networks technique: A review

Polymerization process can be classified as a nonlinear type process since it exhibits a dynamic behaviour throughout the process. Therefore, it is highly complicated to obtain an accurate mechanistic model from the nonlinear process. This predicament always been a “wall” to researchers to be able to devise an optimal process model and control scheme for such a system. Neural networks have succeeded the other modelling and control methods especially in coping with nonlinear process due to their very conciliate characteristics. These characteristics are further explained in this work. The predicament that is encountered by researchers nowadays is lack of data which consequently lead to an imprecise mechanistic model that scarcely conforms to the desired process. The implementations of the neural network model not only restrict to polymerization reactor but to other difficult‐to‐measure parameters such as polymer quality, polymer melts index and mixture of initiators. This work is aimed to manifest ascendancy of neural networks in modelling and control of polymerization process.     

Application of recurrent neural networks in batch reactors: Part I. NARMA modelling of the dynamic behaviour of the heat transfer fluid temperature

This paper is focused on the development of nonlinear models, using artificial neural networks, able to provide appropriate predictions when acting as process simulators. The dynamic behaviour of the heat transfer fluid temperature in a jacketed chemical reactor has been selected as a case study. Different structures of NARMA (Non-linear ARMA) models have been studied. The experimental results have allowed to carry out a comparison between the different neural approaches and a first-principles model. The best neural results are obtained using a parallel model structure based on a recurrent neural network architecture, which guarantees better dynamic approximations than currently employed neural models. The results suggest that parallel models built up with recurrent networks can be seen as an alternative to phenomenological models for simulating the dynamic behaviour of the heating/cooling circuits which change from batch installation to installation.     

Inferential Estimation of Polymer Melt Index Using Sequentially Trained Bootstrap Aggregated Neural Networks

The inferential estimation of a polymer melt index in an industrial polymerization process using aggregated neural networks is presented in this paper. The difficult‐to‐measure polymer melt index is estimated from easy‐to‐measure process variables, and their relationship is estimated using aggregated neural networks. The individual networks are trained on bootstrap re‐samples of the original training data by a sequential training algorithm. In this training method, individual networks, within a bootstrap aggregated neural network model, are trained sequentially. The first network is trained to minimize its prediction error on the training data. In the training of subsequent networks, the training objective is not only to minimize the individual networks' prediction errors but also to minimize the correlation among the individual networks. Training is terminated when the aggregated network prediction performance on the training and testing data cannot be further improved. Application to real industrial data demonstrates that the polymer melt index can be successfully estimated using an aggregated neural network.     

Dynamic On‐Line Reoptimization Control of a Batch MMA Polymerization Reactor Using Hybrid Neural Network Models

A hybrid neural network model based on‐line reoptimization control strategy is developed for a batch polymerization reactor. To address the difficulties in batch polymerization reactor modeling, the hybrid neural network model contains a simplified mechanistic model covering material balance assuming perfect temperature control, and recurrent neural networks modeling the residuals of the simplified mechanistic model due to imperfect temperature control. This hybrid neural network model is used to calculate the optimal control policy. A difficulty in the optimal control of batch polymerization reactors is that the optimization effort can be seriously hampered by unknown disturbances such as reactive impurities and reactor fouling. With the presence of an unknown amount of reactive impurities, the off‐line calculated optimal control profile will be no longer optimal. To address this issue, a strategy combining on‐line reactive impurity estimation and on‐line reoptimization is proposed in this paper. The amount of reactive impurities is estimated on‐line during the early stage of a batch by using a neural network based inverse model. Based on the estimated amount of reactive impurities, on‐line reoptimization is then applied to calculate the optimal reactor temperature profile for the remaining time period of the batch reactor operation. This approach is illustrated on the optimization control of a simulated batch methyl methacrylate polymerization process.     

聚合反应过程质量指标的推理估计混合模型

针对聚合反应过程的非线性、时变性和不确定性，提出了一种多类型混联混合推理估计模型。该模型以过程机理知识为基础框架，以各种神经网络和回归辩识模型的计算结果作为混合模型中各子模型或机理模型的过程参数。为了体现过程的多模式集成特点，该混合模型充分利用各种类型模型的不同特性，既保证按照动力学规律描述聚合反应过程特性，又充分利用现场运行和分析的数据，辩识模型结构参数，使所建模型不必完全依赖对过程特性的认识。将该混合模型用于聚丙烯腈生产过程质量指标的推理估计，现场应用效果证明了这种模型的优良性能。     

Melt index prediction by RBF neural network optimized with an adaptive new ant colony optimization algorithm

Melt index (MI) is a crucial indicator in determining the product specifications and grades of polypropylene (PP). The prediction of MI, which is important in quality control of the PP polymerization process, is studied in this work. Based on RBF (radial basis function) neural network, a soft‐sensor model (RBF model) of the PP process is developed to infer the MI of PP from a bunch of process variables. Considering that the PP process is too complicated for the RBF neural network with a general set of parameters, a new ant colony optimization (ACO) algorithm, N‐ACO, and its adaptive version, A‐N‐ACO, which aim at continuous optimizing problems are proposed to optimize the structure parameters of the RBF neural network, respectively, and the structure‐best models, N‐ACO‐RBF model and A‐N‐ACO‐RBF model for the MI prediction of propylene polymerization process, are presented then. Based on the data from a real PP production plant, a detailed comparison research among the models is carried out. The research results confirm the prediction accuracy of the models and also prove the effectiveness of proposed N‐ACO and A‐N‐ACO optimization approaches in solving continuous optimizing problem. © 2010 Wiley Periodicals, Inc. J Appl Polym Sci, 2010     

Inferential sensors for estimation of polymer quality parameters: Industrial application of a PLS-based soft sensor for a LDPE plant

Low-density polyethylene (LDPE) and ethylene vinyl acetate (EVA) copolymers are produced in free radical polymerization using reactors at extremely high pressure. The reactors require constant monitoring and control in order to minimize undesirable process excursions and meet stringent product specifications. In industrial settings, polymer quality is mainly specified in terms of melt flow index (MI) and density. These properties are difficult to measure and usually unavailable in real time, which leads to major difficulty in controlling product quality in polymerization processes. Researchers have attempted first principles modeling of polymerization processes to estimate end use properties. However, development of detailed first principles model for free radical polymerization is not a trivial task. The difficulties involved are the large number of complex and simultaneous reactions and the need to estimate a large number of kinetic parameters. To overcome these difficulties, some researchers considered empirical neural network models as an alternative. However, neural network models provide no physical insight about the underlying process. We consider data-based multivariate regression methods as alternative solution to the problem. In this paper, some recent developments in modeling polymer quality parameters are reviewed, with emphasis given to the free radical polymerization process. We present an application of PLS to build a soft-sensor to predict melt flow index using routinely measured process variables. Issues of data acquisition and preprocessing for real industrial data are discussed. The study was conducted using data collected form an industrial autoclave reactor, which produces LDPE and EVA copolymer using free radical polymerization. The results indicated that melt index (MI) can be successfully predicted using this relatively straightforward statistical tool.     

Optimal control of batch processes using particle swam optimisation with stacked neural network models

An optimal control strategy for batch processes using particle swam optimisation (PSO) and stacked neural networks is presented in this paper. Stacked neural network models are developed form historical process operation data. Stacked neural networks are used to improve model generalisation capability, as well as provide model prediction confidence bounds. In order to improve the reliability of the calculated optimal control policy, an additional term is introduced in the optimisation objective function to penalize wide model prediction confidence bounds. The optimisation problem is solved using PSO, which can cope with multiple local minima and could generally find the global minimum. Application to a simulated fed-batch process demonstrates that the proposed technique is very effective.     

Evaluation of neural network models and quality forecasting based on process time-series data

Complex industrial process modelling is critically important within the context of industrial intelligence. In recent years, soft sensor techniques based on neural networks have become increasingly popular for modelling nonlinear industrial processes. This paper proposes an integrated framework of neural network modelling and evaluation for nonlinear dynamic processes. This framework achieves an integrated solution for modelling, prediction, evaluation, and network structure parameter selection. It can be applied to noisy sensors and dense data in the time domain. The framework's proposed evaluation mechanism employs two novel evaluation metrics, the variational auto-encoder (VAE)-based Kullback–Leibler (KL) divergence metric and the maximum likelihood estimation-based J metric, which both evaluate the model by mining the statistical properties of the residuals. The framework models the dynamic process with a model order based-gated recurrent units (MOb-GRU) neural network and a modified transformer model. Numerical experiments demonstrate that the evaluation mechanism functions properly in scenarios with multiple signal-to-noise ratios and multiple noise statistical properties and that the framework produces accurate modelling results.     

A novel approach to using neural networks to predict the colour of fibre blends

This work is concerned with the colour prediction of viscose fibre blends, comparing two conventional prediction models (the Stearns–Noechel model and the Friele model) and two neural network models. A total of 333 blended samples were prepared from eight primary colours, including two‐, three‐, and four‐colour mixtures. The performance of the prediction models was evaluated using 60 of the 333 blended samples. The other 273 samples were used to train the neural networks. It was found that the performance of both neural networks exceeded the performance of both conventional prediction models. When the neural networks were trained using the 273 training samples, the average CIELAB colour differences (between measured and predicted colour of blends) for the 60 samples in the test set were close to 1.0 for the neural network models. When the number of training samples was reduced to only 100, the performance of the neural networks degraded, but they still gave lower colour differences between measured and predicted colour than the conventional models. The first neural network was a conventional network similar to that which has been used by several other researchers; the second neural network was a novel application of a standard neural network where, rather than using a single network, a set of small neural networks was used, each of which predicted reflectance at a single wavelength. The single‐wavelength neural network was shown to be more robust than the conventional neural network when the number of training examples was small.     

HYBRID NEURAL MODELLING OF FLUIDISED BED DRYING PROCESS

The aim of this study was to investigate the applicability of hybrid neural models in modelling of drying process. A study aimed at extending a neural network mapping was also carried out. In this approach dimensionless numbers (Re, Ar, H/d) were used as inputs to predict the heat transfer coefficient in a fluidised bed drying process. To produce a data set necessary to train the networks, trials of drying different materials in a fluidised bed were carried out. On the basis of this network, a hybrid model describing the process of drying in a fluidised bed dryer was built. Results obtained were compared not only with available experimental data but also with results obtained using other types of models: a pseudo-dynamic neural model and a classical mathematical model. The analysis of results leads to a conclusion that hybrid models constitute a solid alternative method of process modelling.     

一种新的基于误差矢量化的选择性神经网络集成方法及其在高密度聚乙烯串级反应过程中的应用(英文)

Chemical processes are complex, for which traditional neural network models usually can not lead to satisfactory accuracy. Selective neural network ensemble is an effective way to enhance the generalization accuracy of networks, but there are some problems, e.g., lacking of unified definition of diversity among component neural networks and difficult to improve the accuracy by selecting if the diversities of available networks are small. In this study, the output errors of networks are vectorized, the diversity of networks is defined based on the error vectors, and the size of ensemble is analyzed. Then an error vectorization based selective neural network ensemble (EVSNE) is proposed, in which the error vector of each network can offset that of the other networks by training the component networks orderly. Thus the component networks have large diversity. Experiments and comparisons over standard data sets and actual chemical process data set for production of high-density polyethylene demonstrate that EVSNE performs better in generalization ability.     

HYBRID NEURAL MODELLING OF FLUIDISED BED DRYING PROCESS

The aim of this study was to investigate the applicability of hybrid neural models in modelling of drying process. A study aimed at extending a neural network mapping was also carried out. In this approach dimensionless numbers (Re, Ar, H/d) were used as inputs to predict the heat transfer coefficient in a fluidised bed drying process. To produce a data set necessary to train the networks, trials of drying different materials in a fluidised bed were carried out. On the basis of this network, a hybrid model describing the process of drying in a fluidised bed dryer was built. Results obtained were compared not only with available experimental data but also with results obtained using other types of models: a pseudo-dynamic neural model and a classical mathematical model. The analysis of results leads to a conclusion that hybrid models constitute a solid alternative method of process modelling.     

Modelling of a woodchip refiner using artificial neural network

Neural networks can be an attractive alternative to mathematical modelling of complex and poorly understood processes if input/output data can easily be obtained. Woodchip refining falls into this category. The mechanism of the refining process is still being studied and no thorough models have yet been developed. A feed-forward neural network is proposed for modelling of woodchip refiners. The outputs predicted by the neural network are compared with industrial refiner data. It is also shown that a modified neural network structure can be used to optimize refiner operation and product quality. The advantages and disadvantages of neural network model application in simulation and optimization of industrial processes are discussed.     

Batch-to-batch optimal control of a batch polymerisation process based on stacked neural network models

A neural network based batch-to-batch optimal control strategy is proposed in this paper. In order to overcome the difficulty in developing mechanistic models for batch processes, stacked neural network models are developed from process operational data. Stacked neural networks have enhanced model generalisation capability and can also provide model prediction confidence bounds. However, the optimal control policy calculated based on a neural network model may not be optimal when applied to the true process due to model plant mismatches and the presence of unknown disturbances. Due to the repetitive nature of batch processes, it is possible to improve the operation of the next batch using the information of the current and previous batch runs. A batch-to-batch optimal control strategy based on the linearisation of stacked neural network model is proposed in this paper. Applications to a simulated batch polymerisation reactor demonstrate that the proposed method can improve process performance from batch to batch in the presence of model plant mismatches and unknown disturbances.     

HYBRID NEURAL APPROACHES FOR MODELLING DRYING PROCESSES FOR PARTICULATE SOLIDS

This work presents methods for synthesizing drying process models for particulate solids that combine prior knowledge with artificial neural networks. The inclusion of prior knowledge is investigated by developing two applications with the data from two indirect rotary steam dryers. The first application consisted in the modelling of the drying process of soya meal in a batch indirect rotary dryer, The external and internal mass transfer resistances were associated in the hidden layer of the network to linear and sigmoidal nodes, respectively. The second application consisted in the modelling of the drying process of soya meal in a continuos indirect rotary dryer. The model was constructed using the Semi-parametric Design Approach. The model predicts the evolution of solid moisture content and temperature as a function of the solid position in the dryer. The results show that the hybrid model performs better than the pure “ black box” neural network and default models. They also shows that prior knowledge enhances the extrapolation capabilities of a neural network model,