Energy management systems optimise the usage of the energy supply system while accounting for energy procurement conditions, fluctuating demand and feed-in from renewable sources. This requires a mathematical model of each individual component to accurately predict system behaviour. However, often only highly aggregated information about the efficiency of specific components is available, which can lead to errors. A common example is battery systems, where the data sheet only specifies a single value for efficiency, regardless of operation point. This inaccuracy can lead to significant deviations between the actual and calculated state of charge (SOC). However, determining the operation point-dependent efficiency levels is a time-consuming process that typically requires carrying out dedicated tests on the battery in order to obtain the needed data. This paper proposes an approach where a generic surrogate model is fitted to the target component using 2 different parameter identification methods on timeseries data measured during normal operation. The parameterised models are then used to create detailed maps of the efficiency across the entire operating range. For this paper, a vanadium redox flow battery from the “Smart Region Pellworm” research project was used as the investigated target component. As parameter identification methods, a differential evolution algorithm and an artificial neural network utilising long short-term memory layers, convolutional layers, and dense layers were used. Additionally, a map consisting of a single value for the efficiency from the data sheet and a map, created through the established approach of putting the battery through a dedicated experiment protocol, were used as benchmarks. The efficiency maps were then used to predict the battery’s SOC over 24h in 5 min increments, based on historic charge and discharge time series. The accuracy of the efficiency maps was then evaluated through the mean absolute percentage error (MAPE) between the historic and predicted SOC. The evaluation was carried out on data of a set of non-consecutive days that were not included in the parameter identification data. The results show a median MAPE value of the datasheet-based map of 4.43% while the map based on the ANN parameters achieved 2.61%. The efficiency map based on the parameters of the evolutionary algorithm achieved a median MAPE of 2.43%, while the experiment-based map achieved the best performance with 2.12%. It should be noted that the errors of the SOC slowly add up over time, which can cause a significantly larger deviation from the real value at the end of the forecast horizon. The results show a significant improvement in model accuracy through better efficiency mapping. After an initial effort to create a library of generic surrogate models, the proposed approach should also be applicable to various other energy system components, such as different storage technologies or other energy-converting components.