System Modelling Using Altair romAI Tool

Abstract

The purpose of this thesis was to explore the capabilities of Altair romAI as a software tool for system modelling and simulation. Specifically, to use a simple and fully comprehensible system as an example to evaluate the feasibility of using the romAI neural model imitating it and to assess the benefits of such a replacement. The target system was modelled using Modelica components and a training dataset was generated. Based on this data, several neural network models were trained and evaluated on statistical measures such as losses and coefficients of determination. The trained models were validated using time simulation and the quality of the simulation results were compared with the statistical performance of the trained models. Although some models achieved high accuracy rates, comparative modelling revealed significant deviations between predicted and actual system responses, especially outside the trained input ranges. The variability in training results indicate the difficulty in achieving stable romAI model reliability. The results suggest that despite promising flexibility and potential speed advantages, further improvement is needed to ensure accuracy and reliability in practical applications as well as the importance of comprehensive validation of trained romAI models.