Research Area 3: Integration of the operators

Many advanced control solutions have been abandoned after some period of operation or never went into real operation because they were not accepted by the operators. The reasons why the operators are dissatisfied may be that they do not fully understand the proposed solutions (control moves), or that the control moves violate some constraints or preferences that were not incorporated in the numerical optimization problem. Another critical factor is a lack of robustness that causes a need for frequent intervention by specialized staff that is costly or not available (cf. RA1). More generally speaking, the problem is one of “trust in automation”. It has to be avoided that advanced control systems are overused (i.e. the operators do not intervene if the control algorithm does not react appropriately) or disused (switched off when they would provide a benefit).

Generally, the application of complex algorithms as proposed here that cannot be understood intuitively by the operators (and often not even by the engineers that deploy the solutions). This inevitably reduces trust. The avoidance of distrust requires on the one hand that the control algorithms are very robust, i.e. that they provide at least a feasible, if not optimal, solution under almost all circumstances and that the trustworthiness of the results is indicated in an understandable fashion. On the other hand, the results must be presented in an intuitive, trust-building fashion, and the operators must be able to tune the control algorithm (within reasonable limits) and to experiment with the behavior of the plant themselves to support learning of both the operator and the algorithm. Given proper information, the operators learn to understand the plant performance, the plant dynamics and the results of the optimization better. Due to the model adaptation that is developed in RA2, the algorithms may also “learn” from the operators when they are enforcing an “inspection” of operating regions that otherwise would be classified as suboptimal (and in fact may be suboptimal – this is another example of the “dual control” paradigm discussed in the description of RA2).

Our goal in this research area is to reassess the operator interface of model-based controllers for the specific situation of direct online optimizing control and to design and test a prototype interface based upon the principles and proposals discussed in the papers mentioned above. The interface in the context of optimizing control on the one hand poses fewer problems because the problem formulation of the optimization is more direct and intuitive than in an MPC controller where the performance is strongly influenced by a large number of weights in the cost function and different types of constraints and target formulations can be used. On the other hand, the behavior of the optimizer may be even more difficult to understand in the case of optimizing control because it is no longer aiming at tracking a certain set of variables which can more easily be displayed graphically.

In a survey of the research on the way how human operators control complex dynamic processes in the 1960s and 1970s, Umbers (1979)  pointed out that the characteristic feature of successful operators is that they employ a prediction of the future evolution of the controlled plant and of the influence of their actions on this evolution. In contrast, it could not be verified that a deeper physical understanding of the behavior of the plant led to significant improvements. Thus, the interaction with the operators should adequately support this capability of humans to intuitively anticipate future evolutions even without a deep understanding of the reasons. On the one hand, trust must be built by showing that the predictions are reliable and on the other hand it is important to make use of the ability of the operators to detect when this is no longer the case.

In this research area, an interdisciplinary team will be set up consisting of a post-doc with a background in cognitive science and a researcher with knowledge of optimizing control of chemical processes to explore new options for user-interaction in optimization-based process operations. This research will be geared towards a “synergetic” cooperation of an algorithmic control scheme and operators that makes the best possible use of the capabilities of the operators as well as of the model-based optimization schemes.