Integrated Performance Assessment and Model Validation for Model Based Predictive Control Applications
- Type: Conference Presentation
- Conference Type:
AIChE Spring Meeting and Global Congress on Process Safety
- Presentation Date:
March 21, 2010
- Skill Level:
Significant improvement in product quality, energy savings and profitability of operations in current day process industry can largely be attributed to the model-based control. Key component in such model-based control algorithms is fidelity of the process model. However, along with the time, because of variations in the operating conditions, equipment aging, weather conditions etc., there will be a change in process behavior from the time when the models were identified. If the process changes significantly, the dynamic model used in control applications may no longer be adequate and thus the model-based control scheme may result in poor performance. Hence, system identification must be performed not only at the initial commissioning of a control loop, but also as part of subsequent maintenance activities when plant operating conditions have deviated substantially from nominal conditions to enhance the model-based control. Unfortunately, system identification is often the most demanding and time consuming step in the implementation of such a model-based control application in the process industry. Due to the additional requirement of excessive man-hours in performing system identification and associated productivity loss, there is a strong urge to recognize the necessity for process model re-identification if it is mandatory. Poor agreement between model predictions and output data does not necessarily imply that model re-identification is required, because the unmodeled component may not be necessarily from change in process dynamics. Very often time-varying external disturbances that are entering the process may lead to plant model mismatch (PMM). For these reasons, model validation is an important tool which is used not only as a final check before controller deployment, but also as a monitoring tool while the controller is active to detect any changes in process behavior.
Current work is targeted to develop a simple model validation/ invalidation technique that can be employed in the performance assessment of industrial model based control applications. In this work, a detailed study on how the PMM propagates into the future model predictions as a result of different types of changes in the process dynamics is presented both in open-loop and closed-loop model-based applications. The effects of process changes on various model based applications and their identifiability are also discussed. Then a change detection and model validation algorithm for dynamical systems using a univariate residual cross-correlation approach to test the independence of two time series is presented. The effectiveness and robustness of the proposed algorithm in the presence of time-variant disturbance dynamics are illustrated through Monte-Carlo simulations.