Feedforward and feedback: Multivariable control, like single-loop control, can be accomplished primarily with feedback control and selective (not wholesale) use of feedforward.
Learning Objectives
- Less than 5% of all installed control loops throughout industry take advantage of feedforward capability.
- Feedforward, for all its virtue, adds cost, risk and maintenance to a control loop.
- Model-based multivariable control (MPC) can be thought of as feedforward control on steroids – it employs feedforward models for every matrix location.
All control engineers learn about feedforward early on in their education and careers. It is not complicated and has the potential to reject process disturbances seamlessly, such that there is no impact on the controlled variable. Not even the best-tuned feedback controllers can do that.
In feedforward control, a process disturbance is measured and translated into a change in controller output, which is implemented in phase with the disturbance, so the output rejects the disturbance and the controlled variable continues on unaffected, essentially oblivious to this behind-the-scenes help.
Without feedforward, the same process disturbance would upset the controlled variable, resulting in a process deviation (error) from setpoint, and requiring feedback control action to reject the disturbance over time, usually using the well-known proportional-integral-derivative (PID) algorithm. This is the inherent limitation of feedback: It requires process error to work; feedforward has the potential to prevent error in the first place.
Feedforward is a powerful tool and a fundamental process control concept that almost every process controller in the process industry supports. However, less than 5% of all installed control loops take advantage of feedforward capability. Why is such a powerful and available tool used so sparingly?
Process industry values reliability: Feedforward adds risk
The experience of feedforward in industry over the past 50 to 75 years tells us that industrial process operation places very high value on reliability. Feedforward, for all its virtue, adds cost, risk and maintenance to a control loop. It adds cost because feedforward inputs and models (output translations) have to be provided. It adds risk because performance becomes dependent on the reliability of the additional inputs and models, with models in particular having proven to be generally problematic and unreliable.
Feedforward adds maintenance of the inputs and the models, and performance issues become more difficult to troubleshoot and diagnose, compared to a PID loop by itself. Meanwhile, PID feedback control performance, without feedforward, is almost always considered satisfactory despite inherent transient error.
The balance of these trade-offs over time has meant feedforward has been adopted sparingly and with great discrimination by industry. To qualify for feedforward, the inputs must be reliable, the models must be reliable and it must add significant value. In the whole process industry, only a single application routinely meets these criteria: The textbook case of three-element boiler drum level control. In this application, the input is a routine steam flow measurement; the model is eminently robust (it has a fixed gain of 1.0 and instantaneous dynamics, so there is no potential for either gain or dynamic error); and the added value is high, especially as it mitigates the difficult boiler drum “shrink and swell” effects in this critical loop.
With feedforward control, the potential risk exists, if the feedforward model dynamics are not correct, to double the disturbance, rather than to cancel it – or worse if the model gain is also not accurate. People often use a reduced feedforward gain to hedge this concern, however, there is no substitute for accurate and reliable model dynamics.
History repeats itself: Model-based multivariable control
The experience of model-based multivariable control (MPC) since the 1980s has emphasized these historical lessons of single-loop feedforward. MPC can be thought of as feedforward control on steroids – it employs feedforward models for every matrix location. Where process industry experience traditionally uses less than 5% feedforward models, MPC uses 100%.
Like single-loop feedforward before it, expectations were initially high that MPC would solve process control performance reliability once and for all. Instead, it took support and maintenance to levels not previously dreamed of. The process industry continues to struggle with this today, finding MPC support and maintenance levels nearly unsustainable, but unwilling to give up on the promise of closed loop multivariable control. Fortunately, the process industry is realizing that multivariable control, just like single-loop control, can be reliably accomplished primarily with feedback control and selective (not wholesale) use of feedforward.
Allan Kern, P.E., is owner, APC Performance LLC. Edited by Mark T. Hoske, content manager, Control Engineering, CFE Media, [email protected].
KEYWORDS: Feedforward, feedback, multivariable control
Less than 5% of all installed control loops throughout industry take advantage of feedforward capability.
Feedforward, for all its virtue, adds cost, risk and maintenance to a control loop.
Model-based multivariable control (MPC) can be thought of as feedforward control on steroids – it employs feedforward models for every matrix location.
CONSIDER THIS
What are your control loops doing right now?
