The invention relates to a computer-implemented method for training a predefined system model (MDL) of a parametrized system (SYS), wherein scenarios (SCN) are a parameter set defining a system (SYS) state and/or system (SYS) operation, the method comprising:
(a) providing said system model (MDL),
(b) providing a first set (IST) of scenarios (SCN),
(c) selecting a sub-set (SST) of scenarios (SCN),
(d) acquiring system (SYS) test data (TDT) for the sub-set (SST) of scenarios (SCN),
(e) generating system model (MDL) data for the sub-set (SST) of scenarios (SCN),
(f) determining a modeling error (MER) of the system model (MDL) by comparing the test data (TDT) with the system model (MDL) data.
To reduce the model training complexity the method according to the invention proposes the additional steps:
(g) generating and/or training an error prediction model (EPM) on basis of the modeling errors (MER) determined,
(h) determining an error prediction (ERP) by the error prediction model (EPM) for the first set (IST) of scenarios (SCN),
(i) selecting a certain portion of the first set (IST) of scenarios (SCN) with the highest error prediction (ERP) determined by the error prediction model (EPM) as top error scenarios (TES),
(j) acquiring system (SYS) test data (TDT) for the selected top error scenarios (TES),
(k) training said system model (MDL) with the acquired system (SYS) test data (TDT) for the selected top error scenarios (TES).