Zero-Shot Parameter Estimation of Modelica Models usingPatch Transformer Networks
This paper introduces a transformer-based generative network for rapid parameter estimation of Modelica building models using simulation data from a Functional Mock-up Unit (FMU). Utilizing the \texttt{MixedAirCO2} model from the Modelica Buildings library, we simulate a single-zone mixed-air volume with detailed thermal and \cotwo dynamics. By varying eight physical parameters and randomizing occupancy profiles across 100 simulated systems, we generate a comprehensive dataset. The transformer encoder, informed by room temperature and \cotwo concentration outputs, predicts the underlying physical parameters with high accuracy and without re-tuning (hence, ``zero-shot''). This approach eliminates the need for iterative optimization or can be used to warm-start such optimization-based approaches, enabling real-time control, monitoring, and fault detection in FMU-based workflows.
Ankush Chakrabarty is a Principal Research Scientist at Mitsubishi Electric Research Laboratories (MERL), working at the intersection of machine learning and automatic control—especially Bayesian optimization, meta-learning, and time-series modeling for real-world systems. His recent work spans sample-efficient control and calibration for energy systems, along with AI-driven scenario generation for closed-loop validation. Before MERL, he was a postdoctoral fellow at Harvard, and he earned his Ph.D. in Electrical & Computer Engineering from Purdue University.
Add your biography.