Model Predictive Control -- Where Have We Been and Where Are We Going -- Part 3

About the Authors

Greg McMillan and Stan Weiner bring their wits and more than 80 years of process control experience to bear on your questions, comments and problems. Write to them at [email protected]. Follow McMillan's Control Talk Blog.Stan: In this multipart part series, Mark Darby of CMiD Solutions will share his thoughts on the scope of model predictive control (MPC) applications. The focus in Part 3 is on some practical considerations in MPC setup, maintenance and improvement to meet economic objectives.

Greg: Measurement, analyzer and final control element resolution and threshold sensitivity limits can cause limit cycles. In a PID we can set an integral dead band to turn off integral action when the process variable (PV) is within the limit cycle amplitude of the setpoint (SP). Also we can use external reset feedback of actual valve position to prevent the PID output from changing until the valve moves. How do you prevent the MPC from responding to these cycles?

Mark: Valve stiction is a frequent challenge. In the pretest phase of work, we need to identify and resolve control valves with excessive stiction; if necessary, delaying the project or portions of it. The problem is that the effect of a valve limit cycle can appear in multiple controlled variables. Consider when a cycle is present in a variable at a constraint limit. This will translate into cycles in multiple MVs. Detuning and the use of dead bands are not usually effective. If stiction is present, but is not extreme, there are techniques that are sometimes helpful. This includes more sophisticated filtering of either the measurement or the model bias filter. Also, some MPC controllers have a feature which enforces a minimum move limit for a manipulated variable, which can be set to correspond to the amount of stiction present (otherwise the move is accumulated). If MPC is configured to directly manipulate a valve signal, and the actual valve feedback is available, it should be used to update the model bias. This would be similar to external reset feedback. In testing, it is important that the change to a MV result in a valve change greater than the stiction value; otherwise this will introduce error into the identification results.

Greg: For cascade control with PID, we use external reset feedback of the secondary loop PV to prevent the primary loop from driving the secondary loop SP faster than the secondary can respond causing a burst of oscillations. In the cascade control of an MPC to a PID or even another MPC, how do you prevent driving the secondary PID or MPC faster than it can respond?

Mark: Move suppression, reference trajectory time constant and the length of the control horizon affect how fast or aggressively MPC will adjust manipulated variable PID setpoints. A point to make here is that the response of the PID loops is embedded in the MPC models, which is reflected in the future move plan calculated by the MPC. One can introduce model mismatch if a PID loop is retuned to achieve a significantly different response.

Stan: How do you help the MPC achieve the most benefits?

Mark: You need to understand how planning and scheduling people establish targets for the plant and make sure the MPC is in sync and consistent with them. This requires good communication and documentation. A key issue is what degrees of freedom to give an MPC. For example, should it maximize feed rate? If so, good communication and coordination is required, since it can impact logistical issues associated with feed stock and product volumes. In configuring the MPC, it is critical to get the economics right so it pushes the correct constraint sets. Some limits may depend upon the time of the year based on product demand. An example is a shift from gasoline to diesel in winter. While this is not a huge issue, the completeness and accuracy of this task must not be overlooked.

Greg: We tended to use small MPC focused on a unit operation, whereas in the literature, you see the use of huge and even plant-area wide MPC. We found the MPC to be easier to develop and maintain with time horizons more tailored to the unit operation. We also we did not have to worry about the washout of the effect of process inputs downstream by the filtering effect of volumes. The washout caused an extremely low process gain for downstream CV, resulting in a low condition number for the matrix. In some cases, the model had the wrong sign because of interrelationships. We used disturbance variables between MPC and, if necessary, a cascade of MPC to MPC to increase MPC scope. What are the advantages of increasing the MPC size?

Mark: There are different opinions in the industry on this issue of controller size. Some of this depends on the process objectives as well as the number and location of the constraints. You see differences, for example, between refining and specialty chemicals. It can also depend on the specific MPC controller and the specific MPC practitioner. Bigger is better for the LP or QP in terms of maximizing profit when constraints are present. Dynamic coordination is another reason. Consider two controllers, where a particular MV in an upstream controller is a FF in a downstream controller. The downstream controller would only consider the current value of the FF when it calculated its future MV move plan. If the two controllers are combined, the MVs in the downstream would automatically be coordinated with the upstream MV in the calculation of the future move plan. This trend to bigger controllers began in the 1990s with the increases in computer power and better algorithms. One of the lessons learned is that bigger controllers are more difficult to implement and maintain, as there is a higher requirement of model consistency—for example, the importance of accurate gain ratios, as mentioned earlier. Depending where constraints are in the plant and whether they are plant-limiting can help decide on the number of MPCs. With the configuration capability of the newer tools, one can also start small and grow/combine controllers over time.

Greg: Comment on operator involvement.

Mark: Operator involvement is critical to the success of an MPC project, beginning the day the project team shows up at the plant. Operators are the best source for knowing how the plant operates and its associated challenges. The operators need to be convinced that the controller will help them, not cause problems. We want them to become vested in the long-term success of the MPC.

Stan: How do you make operator involvement more effective?

Mark: Part is training and part is operator goodwill and communication. The operator needs to understand the "what and why" of the actions taken by the MPC. Operator graphics that include past and future trends of predictions can help understanding by showing where the controller has been and where it is going. Active constraint indications, which are common in MPC displays, also help. Today there are better tools for detailing the "why" behind MPC decisions. Operator feedback is important to ensuring the long-term success of the MPC. There may be new situations that demand a better operational understanding or a MPC modification.

Greg: What about the setting of limits?

Mark: Constraint limit setting is an important issue that needs to be consistent across all operating shifts. Setting constraint limits too narrowly will reduce MPC benefits or hide controller problems. As a result, there has been a tendency towards reducing the limits that the operator can change. In this way, if an MPC is not performing well, it will be switched off and force corrective action. Periodic review of limits by operations, planning/scheduling, and control personnel is an important activity.

Stan: What types of MPC metrics do you find useful?

Mark: Many MPCs are on-stream more than 90% of the time. You need something more than just service factor. The real question is MV utilization and how much time is spent at MV limits versus CV limits. An MV pegged at its limit is not producing an economic return. It is also important to monitor the active constraint sets. Operation against new set of constraints should be investigated, as it could be indicative of incorrect limit setting or an actual process change, or a problem in the process or instrumentation. A trend of the bias correction or the uncorrected prediction can provide an indication of model accuracy. Ultimately, you would like the MPC tools to provide diagnostics for MPC performance problems. This is another active development area for MPC technology suppliers.

Greg: Online process metrics have been misleading in some cases, exhibiting inverse response and deceptive short-term inefficiencies. It appears synchronization of process inputs with process outputs may be needed for metrics such as yield and energy efficiency during transitions and upsets because of process dynamics. The metrics may also need some filtering to deal with synchronization errors and noise. What do you see in terms of effectiveness of process metrics?

Mark: As new steady states are achieved, synchronization issues disappear. Online process metrics are particularly effective when they show improvement in plant performance compared to operation before the new or modified MPC went online. This is relatively straightforward when the controller is operating with the same economic objectives. It is a challenge when the objectives change, as the question becomes what to compare current operation against. One approach is to develop an economic model with assumptions for how the process would be controlled without MPC (e.g., the proximity to constraint limits). This is an area that could benefit from standardized approaches.

Greg: We conclude this series with a wrap-up of advanced control myths.

Myth 5 – You can skip the step (bump) tests. Bumps tests are essential for identifying instrumentation and regulatory loop tuning problems. The tests also provide a clearer view for the process engineer and MPC engineer to see what was real, unexpected and strange. As a minimum, a step in each direction should be made at each operating point. Steady-state process simulations can help identify gains for changes in operating conditions and help confirm test results. The old rule is true: If you can see the model from a trend, it is there. Sometimes, the brain can estimate the process gain and dead time better than software.

Myth 6 – You need to completely know your process before you start a MPC application. This would be nice, but often the benefits provided by a model stem from the knowledge discovered during the systematic building and identification procedures. Frequently, the understanding gained by developing models leads to immediate benefits in terms of better setpoints, instrumentation, and valves. The commissioning of the MPC locks in the benefits for varying plant conditions.

Myth 7 – Optimization by pushing constraints will decrease on-stream time. The converse is true. MPC will recognize future violations for unforeseen problems and will back off from the edge to increase on-stream time.

Greg: And a bonus item: This excellent paper on a small MPC application by one of the ISA Mentor program protégés.