Abstract

Recursive data-based modelling is needed for making decision online in varying operating conditions. Recursive algorithms are useful in adapting the parameters within selected memory horizons. Abrupt changes can be handled when the situation change is approved to be drastic. The nonlinear scaling based on generalized norms includes additional alternatives: the norm orders adapt to the gradually changing operating conditions. The drastic shape changes of the scaling functions require full analyses of the orders. The orders can also be stored for different situations and re-used later. Fuzzy inequalities are useful in finding out if the feasible ranges of the most recent period are different from the current active ranges or similar with some of previous feasible ranges. Machine learning is integrated in the system in three levels: (1) finding the appropriate time windows, (2) interactions of feasible levels, and (3) finding decision support when some of feasible ranges need to change. These decisions are supported by expert knowledge. Other model parameters can be included in the analysis. The solution has been tested with measurement data from several application cases. The recursive approach is beneficial in the control and maintenance in varying operating conditions.