Online Model Maintenance via Output Modifier Adaptation

Abstract

This paper describes a method for selecting and adapting the model structure online while running together with a real-time optimization algorithm using Output Modifier Adaptation. The method chooses, among several knowledge-based models, the model structure that is most consistent with the current process data. By allowing the model to change over time, its structure and complexity are able to adapt to the plant data. Competing models are compared based on the modifiers, which are also used by the Output Modifier Adaptation to drive the system to the optimal operating point. The approach is demonstrated on two case studies: a continuous stirred tank reactor and a gas-lifted oil well network. In both cases, the best model structure is chosen among several candidates, and the plant optimum is reached without constraint violations. The case studies indicate that even with a significant amount of noise the modifiers are good indicators for which competing model structures to choose.

Simultaneous online model identification and production optimization using modifier adaptation

“…As such, depending on a weight trading off exploitation and exploration, the criterion for optimization of the control also accounts for the information that will be gained. Based on two case studies in the context of a chemical reactor and subsea oil and gas exploration, we show that our algorithms yield an economic performance comparable to existing methods [5] with improved identification of changes to degraded states.…”

“…Along the same lines, [5,13] designed a MA variant for cases where several models of the plant are available but only one is used at a time. Here, the cost-minimising control input is computed while identifying the model that best fits the plant from an available set.…”

“…As mentioned, the closest work on the joint identification and control input optimization problem addressed in this paper appears in [5,13]. While modifier adaptation is exploited to find the optimal control, identification is based on a heuristic of choosing the least norm defined on the modifiers across the set of models.…”

“…A key technique in order to optimize controls with uncertain knowledge of the plant steady-state is modifier adaptation. It has also been observed in [5] that modifier adaptation can be utilized to perform model identification. In this section, we overview modifier adaptation and the approach in [5].…”

Output modifier adaptation with filter-based constraints

“…A second issue is thus that each submodel being an approximation of the corresponding subplant, their interconnection typically leads to uncertainty propagation when the whole network is simulated, which can, in turn, lead to poorer predictions of the behavior of the plant and thus to even weaker decisions. Several recent extensions to MA and MAy mitigate this second issue by using more plant measurements than necessary [33,34,35,36].…”