“People totally underestimate the effort needed to keep the systems up to date,” says Shell’s Kimpton. “Half of my time is spent maintaining systems. You have to have a close liaison on a daily basis with the panel men to see how the model is doing. You must check that the model is in tune.”
“No model is accurate forever,” agrees Dimitratos. So, he’d like to see tools to identify when a model needs updating. “Every MPC vendor should be thinking of metrics for monitoring the health of the model.”
Invensys, for one, now builds technology for automatically adapting a model right into its base platform, says Martin.
There’s also room for improvement in the building of the original model. It still requires too much work, says Alloway. He’d like to see it done fairly automatically by the DCS. “Vendors can help by providing engineering tools that are easier, faster, more reliable and with better analysis capabilities to facilitate model development,” says Dimitratos. “There’s plenty of room for improvement in the user-friendliness of the engineering tools.”
Another area where vendors could help more is in the commissioning of the controllers, notes Dimitratos. Interfacing the model and the DCS does not happen automatically, but takes significant effort. Vendors should provide an engineering tool to facilitate system integration by users, he says.
A different model
According to Dimitratos, the latest advance in MPC is the emergence of state space models for controllers to replace the empirical models conventionally used. In state space models, every variable has a physical meaning. Plus, the model better matches reality, giving better estimates and leading to better — sometimes significantly better — control actions.
“This is not an incremental improvement,” Dimitratos says. The models will require less engineering effort to develop applications, thereby cutting costs, and will be easier to use and maintain, he says. However, in general, the approach will not open up new opportunities for MPC, he adds.
ABB, for one, now offers a controller using a state space model. By replacing the typical single-input/single-output step response model with a multiple-input/multiple-output state space model, its Optimize IT Predict & Control is said to more accurately predict the effect of disturbances.
AspenTech has deployed a reduced form of a state space model for MPC at a couple of test sites, says Bonner, and will introduce it in the Advanced Control Platform the firm hopes to release in a couple of months. The model trades some of the flexibility of the firm’s rigorous dynamic state space models for speed, so it is fast enough to be used for control. Potential uses include prediction of heat exchanger fouling factors and catalyst aging.
However, Dimitratos sees significant barriers ahead for state space models. “The overwhelming majority of models are not state space. Practitioners are comfortable with other types of models. So, a switch will require a culture change.” In addition, sometimes the benefits will not provide a sufficient financial driver, he adds.
The issue of integration
Farrukh Butt, manager, controls and automation, in the Somerset, N.J., office of contractor Lockwood Greene, voices a widely held concern about current, piecemeal efforts by vendors. “Everybody does their little piece. The issue is putting it all together to have a functioning system.”
The solution will not be for one vendor to do everything, says Dow’s Gipson. “My No. 1 wish is for suppliers to understand what they do well, but realize where others can contribute. It is inevitable that no one supplier can do it all. I’m starting to see an understanding of that by vendors. They are starting to partner together more and more because the competitive environment is forcing this.”
Overall, we can safely predict more and more-capable predictive maintenance and control at more plants.