(243e) Model-Based Output Feedback Control of Networked Process Systems with Event-Triggered Parameter Re-Identification

The design of distributed and supervisory feedback control systems for process networks has been the subject of significant research work in process control (e.g., [1]-[3]). In light of the growing reliance on networked sensor and control systems in plant operations and the increased emphasis placed on smart plant operations [4], the management and optimization of information flow and communication between the plant subsystems are becoming increasingly important considerations in the control problem formulation.

In previous works [5]-[6], the integration of communication and control was addressed in the context of plant-wide control, where a quasi-decentralized model-based networked control structure that enforces closed-loop stability with minimal communication was developed. A set of predictive models were embedded within each local control system, in conjunction with the local state measurements, to generate the local control action at times when communication between the plant subsystems was suspended, and the states of the models were updated when communication was permitted at discrete times. In doing so, a minimum communication rate that guarantees closed-loop stability could be determined. It was also shown that the communication rate depends strongly on the size of the plant-model mismatch, with a larger mismatch requiring a higher rate of information transfer between the plant subsystems.

While a model-based control strategy is an appealing solution that helps reduce the control system’s reliance on the communication medium and its vulnerability to sensor failures and communication outages, the use of a single model with fixed parameters may limit the achievable savings in network resource utilization, especially when plant operating conditions experience variations and drift over time. Such variations, which can be the result of fouling in heat exchangers or catalyst deactivation in catalytic reactors, are unavoidable and invariably lead to process parameters variations that exacerbate the mismatch between the existing model and the actual plant, ultimately leading to increased communication levels to maintain closed-loop stability.

One way to address this problem is by constantly updating the model parameters using data-based model identification techniques so as to try to keep plant-model mismatch to a minimum. This approach was pursued in [7] where a methodology for the integration of model-based control and model identification for networked process systems subject to communication constraints was developed. The methodology aimed to enhance the stability and performance properties of the networked closed-loop system by incorporating within the time-triggered model-based control strategy, a process monitoring scheme that triggers model re-identification whenever a certain instability alarm threshold is breached.

The implementation of this approach, however, assumes the availability of full-state measurements which are used for both controller implementation and model re-identification. In many practical applications, the full-state is not accessible for measurement, and only a finite number of output measurements are available. This situation typically arises due to the difficulty in measuring certain physical variables in real-time such as concentrations of reactive intermediates. The lack of full-state measurements imposes limitations on the implementation of full-state model-based feedback control as well as the data-based identification of model parameters that need to be addressed.

Motivated by these considerations, we present in this work a methodological framework for the integration of model-based output feedback control and model identification for networked process systems subject to limited output measurements, process parameter variations and communication constraints. The framework aims to maintain closed-loop stability in the presence of plant-model mismatch while simultaneously minimizing unnecessary cross communication between the plant subsystems.

Initially, an observer-based output feedback controller is designed for each subsystem. The observer uses measurements of the local output, together with estimates of the outputs of the other subsystems, to compute the control action when communication is suspended. The output estimates are generated from a set of locally-embedded predictive models whose states are updated when communication between the plant subsystems is allowed. Closed-loop stability is then analyzed to determine the maximum allowable update period that can be used to operate the plant-wide control system.

To assess the need for updating the model parameters at any given time, an error detection scheme with a time-varying alarm threshold is devised on the basis of the stability properties of the nominal (drift-free) closed-loop system. When the instability threshold is breached due to parameter variations, process operation is temporarily safe-parked to avert instability. This safe-parking is achieved by temporarily increasing the communication rate to counteract the increased levels of plant-model mismatch. In the meantime, the input and output data collected and exchanged between the plant subsystems during the safe-parking period are used to identify a new model of the plant based on subspace identification techniques. The overall networked closed-loop stability region associated with the new model is characterized and used to select an appropriate model state update rate that can restore the communication frequency to its original level. At this point, the process exits the safe-parking zone and the model parameters are updated. Finally, the implementation of the developed methodology is illustrated using a reactor-separator network example.

References:

[1] Jogwar SS, Baldea M, Daoutidis P. Dynamics and Control of Process Networks with Large Energy Recycle. Industrial & Engineering Chemistry Research. 2009; 48:6087–6097.

[2] Christofides PD, Liu J, Munoz de la Pena D. Networked and Distributed Predictive Control. Springer-Verlag. 2011.

[3] Cai X, Tippett MJ, Xie L, Bao J. Fast Distributed MPC Based on Active Set Method. Computers & Chemical Engineering. 2014; 71:158–170.

[4] Christofides PD, Davis JF, El-Farra NH, Clark D, Harris KRD, Gipson JN. Smart Plant Operations: Vision, Progress and Challenges. AIChE Journal. 2007; 53(11):2734–2741.

[5] Sun Y, El-Farra NH. Quasi-Decentralized Model-Based Networked Control of Process Systems. Computers & Chemical Engineering. 2008; 32(9):2016–2029.

[6] Sun Y, El-Farra NH. Resource-Aware Quasi-Decentralized Control of Networked Process Systems over Wireless Sensor Networks. Chemical Engineering Science. 2012; 69:93-106.

[7] Zedan A, Xue D, El-Farra NH. Integrating Model Identification and Model-Based Control of Networked Process Systems. Proceedings of 13th International Symposium on Process Systems Engineering, vol. 44 of Computer Aided Chemical Engineering, pp. 715–720. Elsevier. 2018.