Multivariable Predictive Control (MPC) is the most widely used advanced process control technology in process industries, with more than 5,000 worldwide applications currently in service. A common and challenging problem is that MPC control performance degrades with time due to inevitable changes in the underlying subject process, such as equipment modifications, changes in operating strategy, instrumentation degradation, etc. Such degradation of control performance results in loss of benefits. Among all causes of control performance degradation, the process model's predictive quality is the primary factor in most cases. To sustain good control performance, the model needs be periodically audited and updated.
To address above challenging problem, a series of innovative technologies were emerging and put into APC practices with significant benefits. Such as automated non-invasive closed-loop step-testing, online model quality auditing, online closed-loop model identification and adaptation, and automated data screening and selection techniques. In this presentation, we will introduce the automated data screening, selection and repairing technology, which changed the traditional way to do an APC project, sped up the APC project tremendously and more important, made the Online Adaptive MPC a reality. This piece of technology can also serve many other data-related modeling and online applications, such as statistical analysis, root-cause analysis, process model calibration, data analytics, Big Data, etc.
By using this technique, all available information associated with process data, such as data sample status, High/Low operating limits, individual variable attributes in MPC controller (MV, DV or CV) and PID control loop associations are utilized to screen data, identify bad sections and repair them with interpolations or model predictions if possible. Internally built MISO (Multi-input, Single-output) predictive models are also employed to assist detection of spikes and unknown disturbances. Based on various criteria, “bad data” for specific applications are automatically identified and marked for removal/skip. Special algorithms for data point interconnection and adjustment techniques are provided to guarantee smooth/continuous replacement values. As a results, the usage of plant test data is maximized in model identification and adaptation. Examples will show how many different type of bad data slices can be automatically identified and how can this auto-data-slicing technique help improve model’s gain accuracy as well as RGA stability in a collinear case.
See more of this Group/Topical: Topical 7: 19th Topical Conference on Refinery Processing