Parameter Estimation of Modelica Building Models Using CasADi

Abstract

Predictive control can substantially improve the energy performance of buildings during operation, but it requires a model of the building to be implemented. Gray-box model identification starts from a physics-based model (white-box element) and complements it with measurements from the operation of the building (black-box element). The level of detail of the original model is limited by the optimization problem that needs to be solved when estimating its parameters. Consequently, it is common to heavily simplify building models hindering the intelligibility of their parameters and limiting their application potential. This paper investigates the accuracy and scalability of different transcription methods for parameter estimation of building models. The methodology starts from a Modelica model as an initial guess which is transferred to CasADi using the Functional Mockup Interface to solve the parameter estimation problem. The study demonstrates the high effectiveness of multiple shooting. Single shooting and direct collocation could be more suitable for setups with faster integration times or with increased granularity in the training data, respectively.