Authors:
Rian Beck
;
Sudarsan Venkatesan
;
Joram Meskens
;
Jeroen Willems
;
Edward Kikken
and
Bruno Depraetere
Affiliation:
Flanders Make, Lommel, Belgium
Keyword(s):
Adaptive Control, Clustering, Context-Adaptive Control, Machine Learning in Control Applications.
Abstract:
In this paper we present an approach to adapt the parameters of controllers during operation. It is targeted at industrial adoption, relying on controllers of the same type currently in use, but adjusting their gains at run-time based on varying system and / or environment conditions. As the key contribution of this paper we present a method to discover what condition variations warrant a control adaptation for cases where this is not known up front. The goal is not to achieve a better performance than other adaptive control schemes, but to provide a different method of designing or deciding how to build adaptation logic. To achieve this we use data-driven methods to, in an offline preprocessing step: (I) derive features that quantify system / environment variations, (II) optimize the control parameters for the distinct feature values, (III) search for clusters in the multi-dimensional space of both these features and control parameters, looking for sets of similar features as well a
s control parameters to be used. Once a set of clusters is defined, an online adaptive controller is then synthesized by (I) building a classifier to determine which cluster the currently observed conditions belong to, and (II) selecting the optimal control parameters for that cluster. This paper provides a first illustration of the method, without theoretical analysis, on an example case of energy management for a hybrid electrical vehicle, for which an Equivalent Consumption Minimization Strategy controller is built whose parameters are adjusted as the detected cluster changes. The results show an increase in energy-efficiency of the adaptive control method over the non-adaptive one in a variety of scenarios.
(More)