Efficient Combined Physical and Neural Network Modeling and Simulation Using the Functional Mock-up Interface

This paper presents efficient methods in the context of modeling, optimization and simulation of artificial neural networks (NNs) embedded in Functional Mock-up Unit (FMUs). Machine learning models like NNs are widely used in many applications such as control, prediction and decision making. Their integration into simulation environments, that are for example used in the control engineering community, is still a challenging task. With the Functional Mock-up Interface (FMI), a standardized modeling format for dynamic models exists, which can be used to integrate NNs from popular deep learning frameworks into such simulation environments. In this paper, a method to integrate NNs into FMUs is presented and it is shown how to combine NN-based and physical simulation models in a single FMU. The proposed methods are validated with several application examples and the results emphasize its effectiveness when used within the model-based design process for a wide range of dynamic models.