Online Feedback Optimization of Compressor Stations with Model Adaptation using Gaussian Process Regression

arXiv:2212.03604 · eess.SY, cs.SY · Submitted 2022-12-07 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.

Rosa: I'm Rosa.

Dev: And I'm Dev, here with Taro.

Rosa: We're looking at the paper: "Online Feedback Optimization of Compressor Stations with Model Adaptation using Gaussian Process Regression".

Dev: Here's the gist of the paper. Online Feedback Optimization with model adaptation using Gaussian Process Regression mitigates plant-model mismatch by explicitly improving compressor models online to achieve a significant reduction in power consumption when compared to standard Online Feedback Optimization Problem and Motivation The paper addresses the issue of suboptimal results in Online Feedback Optimization when using a linear plant model due to plant-model mismatch, which is a difference between a mathematical model and a real system The presence of plant-model mismatch leads to suboptimal results when using this approach This work presents a novel application of Online Feedback Optimization with online model adaptation using Gaussian Process regression to overcome this shortcoming Model Adaptation Mechanism The core of the mitigation strategy involves online model adaptation through Gaussian process regression to estimate the difference between real characteristics and the assumed model The error between the real efficiency and the approximation is modeled as a function of mass flow and pressure ratio, which is then treated as a Gaussian process with a squared exponential covariance function This allows for the estimation of performance characteristics for gradient-based optimization The adaptation is performed iteratively by extending the data used for estimation of parameters, which improves the prediction for the error Optimization Formulation The problem is formulated as minimizing power consumption subject to demand constraints using Online Feedback Optimization The optimization problem is defined as min u,y Φ(u, y) subject to y = h(u) and constraints on inputs and outputs The input-output formulation relates the target mass flows to the measured flows and power consumption of each compressor The quadratic optimization problem is solved using quadprog in Matlab Version 9.10.0.1739362 Performance and Results The approach demonstrates that Online Feedback Optimization with model adaptation reduces the power consumption increase caused by imperfect knowledge from 5% to only 0.8% This reduction is achieved by learning the mismatch between the plant and model online, both for parametric and structural mismatch The Gaussian process regression captures the structural mismatch as well as parametric mismatch, leading to better predictions for set-points that have not yet been observed The demand constraint is satisfied by all tested approaches, proving that the reformulation of the equality constraint in (33) is sufficient for satisfaction of the equality constraint (9) The proposed approach successfully mitigates plant-model mismatch while satisfying demand The combination of the new model adaptation approach based on Gaussian process regression with Online Feedback Optimization mitigates the increase in power consumption from 5% to 0.8% The proposed approach is also able to handle both parametric and structural mismatch, while satisfying the required demand Future work will explore the stochastic nature of Gaussian process regression and analyze theoretical properties of Online Feedback Optimization with model adaptation Acknowledgement Financial support from ABB for the Autonomous Industrial Systems Laboratory is gratefully acknowledged References Ahmed, A., Zagorowska, M., del Rio-Chanona, E.A., Mercang¨oz, M., 2022.

Rosa: Application of Gaussian processes to online approximation of compressor maps for load-sharing in a compressor station BSI, 2014.

Dev: BS EN 12583:2014.

Rosa: Gas infrastructure.

Paper summary: Rosa: So to recap, the paper "Online Feedback Optimization of Compressor Stations with Model Adaptation using Gaussian Process Regression" is about overcoming the limitations of standard online feedback optimization when a linear plant model doesn't perfectly match the real system >

Dev: The core idea is to use Gaussian process regression to estimate that difference between the mathematical model and reality, which they call plant-model mismatch >

Rosa: They are treating this error as something that depends on mass flow and pressure ratio, using a squared exponential covariance function to model how that error changes over time >

Dev: The goal is to use this learned mismatch information to improve the predictions for the optimization calculations without having to solve a massive nonlinear constrained optimization problem every single time >

Rosa: They formulate the problem as minimizing power consumption subject to demand constraints, and they show how this adaptation helps reduce the power consumption increase caused by imperfect knowledge from five percent down to only zero point eight percent >

Dev: That reduction is significant because it shows that learning the mismatch online can substantially improve efficiency in these compressor stations >

Taro: The paper tackles both parametric and structural mismatch, which means it's not just about small errors in the parameters, but also how the model is fundamentally set up wrong >

Rosa: They demonstrate that this technique allows for better predictions of set-points even for conditions they haven't seen before because the error estimation gets refined with every new data point collected >

Dev: The paper also confirms that by reformulating how they handle the equality constraint in equation nine they can still satisfy all the demand constraints in their tested scenarios > <ref:2212.03604#pg3>

Taro: It sounds like a solid approach for handling uncertainty in dynamic systems where you're trying to run something continuously, which is exactly what we deal with when we test failure modes >

Rosa: So it seems like the big claim is that this online adaptation method makes the optimization much more reliable and less sensitive to model inaccuracies in real-time operations >

Dev: It's about making sure that even if the system behaves a little differently than predicted, we still get close to the true optimal point for power usage >

Conclusion: Rosa: So looking at "Online Feedback Optimization of Compressor Stations with Model Adaptation using Gaussian Process Regression," this paper is really about improving control in gas compressor stations where you have to balance different compressors based on demand >

Dev: The authors are M. Zagorowskaa, M. Degnerb, L. Ortmannb, A. Ahmeda, S. Bolognanib, and E A Del Rio Chanonaa from Imperial College London and ETH Zurich >

Rosa: What they’re saying simply is that standard online feedback optimization can give you suboptimal results because the mathematical model of the physical compressor system isn't perfect >

Dev: They introduce Gaussian process regression as a tool to learn how much that model differs from the real plant characteristics in real time, based on things like mass flow and pressure ratio >

Rosa: The main implication is that this online adaptation makes the optimization much more reliable for energy efficiency, cutting power consumption increases from five percent down to zero point eight percent >

Dev: It suggests that for complex industrial processes where you can't perfectly model everything beforehand, continuously updating your model based on new measurements is a viable way to keep things running efficiently >

Taro: For someone who just listens to this show, it means that even if your control system has some inherent errors or the physical machine drifts slightly, you can adapt your optimization strategy online without completely failing >

Rosa: It’s about building resilience into the optimization process itself by letting it learn from its own mistakes and discrepancies >

M. Zagorowskaa, M. Degnerb, L. Ortmannb, A. Ahmeda, S. Bolognanib, E. A del Rio Chanonaa, M Mercangóza

Department of Chemical Engineering, Imperial College London · Automatic Control Laboratory, ETH Zurich

eess.SY, cs.SY

Submitted: 2022-12-07

Updated: 2022-12-07

Journal ref: Journal of Process Control, 2023 121, 119 - 133

DOI: 10.1016/j.jprocont.2022.12.001

License: http://creativecommons.org/licenses/by-nc-nd/4.0/

Importance score: 88/100

The gist: Online Feedback Optimization with model adaptation using Gaussian Process Regression mitigates plant-model mismatch by explicitly improving compressor models online to achieve a significant reduction…

Key concepts

Online Feedback Optimization
This is a control strategy where a system continuously monitors its performance and adjusts its operations in real-time to achieve an optimal goal, such as minimizing energy use. It involves solving an optimization problem repeatedly based on the latest measurements of the physical system.
Plant-Model Mismatch
This occurs when the mathematical model used to predict a compressor's behavior does not perfectly match how the real physical compressor actually operates. This discrepancy leads to suboptimal control decisions and increased energy waste if not addressed.
Gaussian Process Regression
This is a machine learning technique used here to estimate the unknown errors between the real system performance and the assumed mathematical model. It models this error as a probability distribution, allowing the system to adapt its predictions iteratively by using new data points.

Terminology

Summary

Online Feedback Optimization with model adaptation using Gaussian Process Regression mitigates plant-model mismatch by explicitly improving compressor models online to achieve a significant reduction in power consumption when compared to standard Online Feedback Optimization

Problem and Motivation

The paper addresses the issue of suboptimal results in Online Feedback Optimization when using a linear plant model due to plant-model mismatch, which is a difference between a mathematical model and a real system The presence of plant-model mismatch leads to suboptimal results when using this approach This work presents a novel application of Online Feedback Optimization with online model adaptation using Gaussian Process regression to overcome this shortcoming

Model Adaptation Mechanism

The core of the mitigation strategy involves online model adaptation through Gaussian process regression to estimate the difference between real characteristics and the assumed model The error between the real efficiency and the approximation is modeled as a function of mass flow and pressure ratio, which is then treated as a Gaussian process with a squared exponential covariance function This allows for the estimation of performance characteristics for gradient-based optimization The adaptation is performed iteratively by extending the data used for estimation of parameters, which improves the prediction for the error

Optimization Formulation

The problem is formulated as minimizing power consumption subject to demand constraints using Online Feedback Optimization The optimization problem is defined as min u,y Φ(u, y) subject to y = h(u) and constraints on inputs and outputs The input-output formulation relates the target mass flows to the measured flows and power consumption of each compressor The quadratic optimization problem is solved using quadprog in Matlab Version 9.10.0.1739362

Performance and Results

The approach demonstrates that Online Feedback Optimization with model adaptation reduces the power consumption increase caused by imperfect knowledge from 5% to only 0.8% This reduction is achieved by learning the mismatch between the plant and model online, both for parametric and structural mismatch The Gaussian process regression captures the structural mismatch as well as parametric mismatch, leading to better predictions for set-points that have not yet been observed The demand constraint is satisfied by all tested approaches, proving that the reformulation of the equality constraint in (33) is sufficient for satisfaction of the equality constraint (9)

Conclusion

The proposed approach successfully mitigates plant-model mismatch while satisfying demand The combination of the new model adaptation approach based on Gaussian process regression with Online Feedback Optimization mitigates the increase in power consumption from 5% to 0.8% The proposed approach is also able to handle both parametric and structural mismatch, while satisfying the required demand Future work will explore the stochastic nature of Gaussian process regression and analyze theoretical properties of Online Feedback Optimization with model adaptation

Acknowledgement

Financial support from ABB for the Autonomous Industrial Systems Laboratory is gratefully acknowledged

References

Ahmed, A., Zagorowska, M. p. 0

Where the paper says it

Mercang¨oza aDepartment of Chemical Engineering, Imperial College London, South Kensington, SW7 2AZ London, UK (m. zagorowska@imperial.

, del Rio-Chanona, E. p. 0
Where the paper says it

ahmed21@imperial. del-rio-chanona@imperial.

A., Mercang¨oz, M. p. 0
Where the paper says it

del Rio Chanonaa , M. Mercang¨oza aDepartment of Chemical Engineering, Imperial College London, South Kensington, SW7 2AZ London, UK (m.

p. 0
Where the paper says it

del Rio Chanonaa , M. Mercang¨oza aDepartment of Chemical Engineering, Imperial College London, South Kensington, SW7 2AZ London, UK (m.

p. 0
Where the paper says it

del Rio Chanonaa , M. Mercang¨oza aDepartment of Chemical Engineering, Imperial College London, South Kensington, SW7 2AZ London, UK (m.

p. 0
Where the paper says it

del Rio Chanonaa , M. Mercang¨oza aDepartment of Chemical Engineering, Imperial College London, South Kensington, SW7 2AZ London, UK (m.

p. 0
Where the paper says it

del Rio Chanonaa , M. Mercang¨oza aDepartment of Chemical Engineering, Imperial College London, South Kensington, SW7 2AZ London, UK (m.

p. 0
Where the paper says it

del Rio Chanonaa , M. Mercang¨oza aDepartment of Chemical Engineering, Imperial College London, South Kensington, SW7 2AZ London, UK (m.

p. 0
Where the paper says it

del Rio Chanonaa , M. Mercang¨oza aDepartment of Chemical Engineering, Imperial College London, South Kensington, SW7 2AZ London, UK (m.

, 2022. Application of Gaussian processes to online approximation of compressor maps for load-sharing in a compressor station

BSI, 2014. BS EN 12583:2014. p. 2

Where the paper says it

Background 1. Compressor stations A compressor station is typically a part of a natural gas transport network and provides a boost for transporting the gas to the receivers (BSI, 2014).

Gas infrastructure. p. 2
Where the paper says it

Background 1. Compressor stations A compressor station is typically a part of a natural gas transport network and provides a boost for transporting the gas to the receivers (BSI, 2014).

Compressor stations - Functional requirements

Chu, F., Dai, B., Lu, N., Ma, X., Wang, F., 2018. Improved fast model migration method for centrifugal compressor based on Bayesian algorithm and Gaussian process model

Cortinovis, A., Ferreau, H.J., Lewandowski, D., Mercang¨oz, M., 2015. Experimental evaluation of MPC-based anti-surge and process control for electric driven centrifugal gas compressors

Cortinovis, A., Mercang¨oz, M., Zovadelli, M., Pareschi, D., De Marco, A., Bittanti, S., 2016. Online performance tracking and load sharing optimization for parallel operation of gas compressors

Degner, M., 2021. Online Feedback Optimization for Gas Compressors

Egeland, O., Gravdahl, J. p. 9

Where the paper says it

The efficiency can then be obtained as: η = Wiso Wact (3) However, gas properties as well as the temperatures may be unknown. Thus, the efficiency is usually approximated as a second order polynomial function of the mass flow through a compressor, mc i and pressure ratio Πi as: ηˆ poly i (mc i , Πi) =α0 + α1mc i + α2Πi + α3mc iΠi + α4(mc i ) 2 + α5(Πi) 2 (4) The polynomial model from (4) is considered a good representation of the functional form of compressor efficiency maps (Egeland and Gravdahl, 2002).

T., 2002. Modeling and simulation for automatic control

Gentsch, M., King, R., 2020. Real-time estimation of a multi-stage centrifugal compressor performance map considering real-gas processes and flexible operation

Hauswirth, A., Bolognani, S., Hug, G., D¨orfler, F., 2021a. Optimization algorithms as robust feedback controllers

Hauswirth, A., Bolognani, S., Hug, G., Dorfler, F., 2021b. Timescale separation in autonomous optimization

Jenicek, T., Kralik, J., et al., 1995. Optimized control of generalized compressor station

Jung, J., Lee, W.J., Park, S., Kim, Y., Lee, C.J., Han, C., 2017. Improved control strategy for fixed-speed compressors in parallel system

Kashani, A.H.A., Molaei, R., 2014. Techno-economical and environmental optimization of natural gas network operation

Korkmaz, B.S., Mercang¨oz, M., 2022. Data driven modelling of centrifugal compressor maps for control and optimization applications

Kumar, S., Cortinovis, A. p. 2

Where the paper says it

(2013), Milosavljevic et al. (2016), and Kumar and Cortinovis (2017).

p. 2
Where the paper says it

(2013), Milosavljevic et al. (2016), and Kumar and Cortinovis (2017).

p. 2
Where the paper says it

(2013), Milosavljevic et al. (2016), and Kumar and Cortinovis (2017).

p. 2
Where the paper says it

(2013), Milosavljevic et al. (2016), and Kumar and Cortinovis (2017).

, 2017. Load sharing optimization for parallel and serial compressor stations

Liptak, B.G., 2005. Instrument Engineers’ Handbook, Volume Two: Process Control and Optimization

Milosavljevic, P., Cortinovis, A., Marchetti, A.G., Faulwasser, T., Mercang¨oz, M., Bonvin, D., 2016. Optimal load sharing of parallel compressors via modifier adaptation

Milosavljevic, P., Marchetti, A.G., Cortinovis, A., Faulwasser, T., Mercang¨oz, M., Bonvin, D., 2020. Real-time optimization of load sharing for gas compressors in the presence of uncertainty

Mokhatab, S., Poe, W.A., Mak, J.Y., 2015. Handbook of Natural Gas Transmission and Processing

Nørstebø, V.S., 2008. Optimum Operation of Gas Export Systems

Ortmann, L., Hauswirth, A. p. 1

Where the paper says it

That said, such machines are typically controlled by embedded units such as programmable logic controllers or so called edge computing devices, which tend to have relatively limited computational capabilities. In this regard, Online Feedback Optimization (OFO) is a model-based method that can steer such a machine or a collection of these machines to an optimum point without solving a demanding nonlinear constrained optimization problem (Hauswirth et al.

, Caduff, I., D¨orfler, F., Bolognani, S., 2020. Experimental validation of feedback optimization in power distribution grids

Osiadacz, A., Bell, D.J., 1981. A local optimization procedure for a gascompressor station

Paparella, F., Dom´ınguez, L., Cortinovis, A., Mercang¨oz, M., Pareschi, D., Bittanti, S., 2013. Load sharing optimization of parallel compressors

Picallo, M., Ortmann, L., Bolognani, S., D¨orfler, F., 2021. Adaptive real-time grid operation via online feedback optimization with sensitivity estimation

Rasmussen, C.E., 2003. Gaussian processes in machine learning

Rasmussen, C.E., Williams, C. p. 3

Where the paper says it

In the current work we mitigate the need for multiple measurements by using Gaussian process regression. Gaussian process regression Gaussian processes regression is a non-parametric generalisation of standard linear regression (Rasmussen and Williams, 2006).

K., 2006. Gaussian processes for machine learning

Ren, L., Zhou, S., Peng, T., Ou, X., 2021. A review of CO2 emissions reduction technologies and low-carbon development in the iron and steel industry focusing on china

Riungu, E.K., Moses, P. p. 1

Where the paper says it

The results are shown for load sharing optimization in a realistic compressor station with centrifugal compressors working in parallel. Industrial motivation There is renewed interest to increase the energy-efficiency of industrial operations as a quick way for reducing the carbon emissions associated with the consumed energy (Riungu and Moses, 2022; Ren et al.

M., 2022. Economic analysis and energy savings of variable speed drives in fans application — a case study Capwell milling factory

Vilalta, P.C., Wan, H., Patnaik, S.S., 2019. Centrifugal compressor performance prediction using Gaussian process regression and artificial neural networks

Wu, S., Rios-Mercado, R.Z., Boyd, E.A., Scott, L.R., 2000. Model relaxations for the fuel cost minimization of steady-state gas pipeline networks

Xenos, D.P., 2015. Optimal Operation of Industrial Compressor Stations in Systems with Large Energy Consumption

Zagorowska, M., Thornhill, N. p. 1

Where the paper says it

In this regard, Online Feedback Optimization (OFO) is a model-based method that can steer such a machine or a collection of these machines to an optimum point without solving a demanding nonlinear constrained optimization problem (Hauswirth et al. The performance of rotating machinery tends to degrade over time and accurate models for such systems are generally not available (Zagorowska and Thornhill, 2020).

F., 2020.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Online Feedback Optimization can steer a process plant to its optimal operating point without explicitly solving a nonlinear constrained optimization problem by leveraging a linear plant model and feedback from measurements. This enables computational efficiency for complex problems with fast dynamics.

  2. Online model adaptation using Gaussian Process regression can estimate the difference between the real characteristics and the assumed model, enabling Online Feedback Optimization to overcome plant-model mismatch. This mitigation reduces power consumption increases from 5% to 0.8% in a compressor station scenario.

  3. The improved AI system can perform load sharing optimization for compressor stations with unknown characteristics by formulating it as an OFO problem that minimizes total power consumption while using online GP regression to estimate performance characteristics for gradient-based optimization.

Abstract

Online Feedback Optimization is a method used to steer the operation of a process plant to its optimal operating point without explicitly solving a nonlinear constrained optimization problem. This is achieved by leveraging a linear plant model and feedback from measurements. However the presence of plant-model mismatch leads to suboptimal results when using this approach. Learning the plant-model mismatch enables Online Feedback Optimization to overcome this shortcoming. In this work we present a novel application of Online Feedback Optimization with online model adaptation using Gaussian Process regression. We demonstrate our approach with a realistic load sharing problem in a compressor station with parametric and structural plant-model mismatch. We assume imperfect knowledge of the compressor maps and design an Online Feedback Optimization controller that minimizes the compressor station power consumption. In the evaluated scenario, imperfect knowledge of the plant leads to a 5% increase in power consumption compared to the case with perfect knowledge. We demonstrate that Online Feedback Optimization with model adaptation reduces this increase to only 0.8%, closely approximating the case of perfect knowledge of the plant, regardless of the type of mismatch.

Sources

Related papers