Process structure-based recurrent neural network modeling for predictive control: A comparative study

Alhajeri, Mohammed S.; Liu, Junwei; Wu, Zhe; Albalawi, Fahad; Christofides, Panagiotis D.

doi:10.1016/j.cherd.2021.12.046

Cited by 41 publications

(14 citation statements)

References 27 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…This is the key reason for exploring the use of sparse identification modeling in the context of MPC. In ref , for a similar process system, i.e., CSTRs modeled via Aspen Plus Dynamics, it was demonstrated that the average time to solve the optimization problem in the MPC took 2.1161 min (127 s), while the proposed dropout-SINDy-based MPC took an average of 42 s. While the system in ref is not identical to the one studied here, the number of control actions to be computed by the MPC and the computational power of the computer were very similar, thereby expecting similar computational efficiency results.…”

Section: Preliminariesmentioning

confidence: 76%

Modeling and Control of Nonlinear Processes Using Sparse Identification: Using Dropout to Handle Noisy Data

Abdullah

Alhajeri

Christofides

2022

Ind. Eng. Chem. Res.

Self Cite

View full text Add to dashboard Cite

Sparse identification of nonlinear dynamics (SINDy) is a recent nonlinear modeling technique that has demonstrated superior performance in modeling complex time-series data in the form of first-order ordinary differential equations (ODEs), which are explicit and continuous in time. However, a crucial step in the SINDy algorithm involves estimating the time derivative of the states from the discrete, measured data. Therefore, the presence of noise can greatly deteriorate the performance if it is not carefully considered and accounted for. In this work, SINDy is used with ensemble learning, where multiple models are identified to improve the overall/final nonlinear model’s performance. Specifically, in the SINDy algorithm, a fraction of the library functions considered for the ODE model representation are randomly dropped out in each submodel to favor model sparsity and stability at the possible risk of lowering the model accuracy. This trade-off is controlled by manipulating the fraction of the library functions dropped out and the total number of models generated, both of which are considered as hyperparameters to be tuned in the proposed algorithm. Data from open-loop simulations of a large-scale chemical plant are generated using the well-known high-fidelity process simulator, Aspen Plus Dynamics, and corrupted with substantial sensor noise to be implemented in the newly proposed algorithm, dropout-SINDy. The dropout-SINDy models obtained from training with the noisy data are then tested in open-loop simulations to demonstrate accurate identification of the steady-state and reasonably close transient behavior under a variety of initial conditions and manipulated input values. Finally, the constructed models are used in a Lyapunov-based model predictive controller to control the large-scale Aspen process, meeting desired closed-loop stability and performance specifications.

show abstract

Section: Preliminariesmentioning

confidence: 76%

Modeling and Control of Nonlinear Processes Using Sparse Identification: Using Dropout to Handle Noisy Data

Abdullah

Alhajeri

Christofides

2022

Ind. Eng. Chem. Res.

Self Cite

View full text Add to dashboard Cite

show abstract

“…Three reactions are taken place in the production process, with one primary reaction and two other side reactions. This process consists of two well‐mixed and nonisothermal continuous stirred tank reactors (CSTRs) that are placed in series as Reference 10. The primary reaction is a second‐order, exothermic, and irreversible reaction as follows:

C_{2} H_{4} goodbreak+ C_{6} H_{6} \to C_{8} H_{10}

In Aspen, a steady‐state model is first constructed to ensure the balance of material and energy, and then the pressure drop and the reactor configuration parameters are specified to ensure the connection between each unit and keep the reactant flow in the process.…”

Section: Application To Chemical Process Examplesmentioning

confidence: 99%

Modeling and predictive control of nonlinear processes using transfer learning method

Xiao

2023

AIChE Journal

Self Cite

View full text Add to dashboard Cite

This work develops a transfer learning (TL) framework for modeling and predictive control of nonlinear systems using recurrent neural networks (RNNs) with the knowledge obtained in modeling one process transferred to another. Specifically, transfer learning uses a pretrained model developed based on a source domain as the starting point, and adapts the model to a target process with similar configurations. The generalization error for TL‐based RNN (TL‐RNN) is first derived to demonstrate the generalization capability on the target process. The theoretical error bound that depends on model capacity and the discrepancy between source and target domains is then utilized to guide the development of pretrained models for improved model transferability. Subsequently, the TL‐RNN model is utilized as the prediction model in model predictive controller (MPC) for the target process. Finally, the simulation study of chemical reactors via Aspen Plus Dynamics is used to demonstrate the benefits of transfer learning.

show abstract

“…However, in the field of modeling systems for mineral processing (grinding, classifying and concentrating), artificial neural networks are relatively recent, but their use is increasing in this type of systems due to the efficiency of the results that they can generate, avoiding the implementation of complex calculations with better performance [16], [24]. Currently, the intelligence systems by means of artificial neural networks can be summarized into four structures: i) supervised learning: the neural network learns a set of inputs and the desired outputs to solve the problem [53]- [56], ii) direct inverse learning: the neural network learns from the feedback of a system, so that, when the signal is obtained, it determines the parameters to be performed [52], [57]- [60], iii) utility backpropagation: this structure optimizes the mathematical equation that represents the system, where its main disadvantage is that it requires a model of the system to be analyzed [61]- [65] and iv) adaptive critical learning: similar to the utility backpropagation structure, but without the need for a model of the plant [66]- [68]. Although this type of structures are present and well accepted in different industrial processes, it is evident that in mineral processing applications and especially in the prediction of variables of interest such as mineral recovery, the existing studies of this type of design are based on simulations, this research being a starting point for the implementation of intelligent systems in gravimetric concentration equipment where experimental data obtained from a pilot scale jig is worked on.…”

Section: Related Workmentioning

confidence: 99%

Estimation of Recovery Percentage in Gravimetric Concentration Processes using an Artificial Neural Network Model

Alarcón¹,

Rivera-M²,

Golondrino³

2022

IJACSA

View full text Add to dashboard Cite

The concentrate process is the most sensitive in mineral processing plants (MPP), and the optimization of the process based on intelligent computational models (machine learning for recovery percentage modelling) can offer significant savings for the plant. Recent theoretical developments have revealed that many of the parameters commonly assumed as constants in gravity concentration modelling have a dynamic nature; however, there still lacks a universal way to model these factors accurately. This paper aims to understand the model effect of operational parameters of a jig (gravimetric concentrator) on the recovery percentage of the interest mineral (gold) through empirical modeling. The recovery percentage of mineral particles in a vibrated bed of big particles is studied by experimental data. The data used for the modelling were from experimental test in a pilot-scale jig supplemented by a twomonth field sampling campaign for collecting 151 tests varying the most significant parameters (amplitude and frequency of pulsation, water flow, height of the artificial porous bed, and particle size). It is found the recovery percentage (%R) decreases with increasing pulsation amplitude (A) and frequency (F) when the size ratio of small to large particles (d/D) is smaller than 0.148. An empirical model was developed through machine learning techniques, specifically an artificial neural network (ANN) model was built and trained to predict the jig recovery percentage as a function of operation parameters and is then used to validate the recovery as a function of vibration conditions. The performance of the ANN model was compared with a new 65 experimental data of the recovery percentage.Results showed that the model (R 2 = 0.9172 and RMSE = 0.105) was accurate and therefore could be efficiently applied to predict the recovery percentage in a jig device.

show abstract

Process structure-based recurrent neural network modeling for predictive control: A comparative study

Cited by 41 publications

References 27 publications

Modeling and Control of Nonlinear Processes Using Sparse Identification: Using Dropout to Handle Noisy Data

Modeling and Control of Nonlinear Processes Using Sparse Identification: Using Dropout to Handle Noisy Data

Modeling and predictive control of nonlinear processes using transfer learning method

Estimation of Recovery Percentage in Gravimetric Concentration Processes using an Artificial Neural Network Model

Contact Info

Product

Resources

About