Now it looks more real - A Study of Metrics and Resolution for the Calibration of Dynamic Simulation

The calibration of dynamic building simulation is an often essential but challenging process. There are approaches to formulate it mathematically as the minimization of a discrepancy function between simulation and measurements. In these approaches, the discrepancy is usually defined based on the mean square error (MSE) between vectors of simulated and measured quantities. The MSE metric reflects the area between the corresponding curves, without accounting for the similarity of curve shapes. Thus, it may overestimate the discrepancy of quantities subject to strong variations, especially in the case of on/off behavior with hysteresis and phase shifts. The aim of this contribution is to investigate whether the MSE still represents an effective metric in such cases, and to examine the usefulness of an alternative metric. This alternative metric is based on dynamic time warping (DTW), which allows similar features to be matched with some distortion on the time axis. The question is elaborated with the example of a waterto- water heat pump, which may cycle on and off at a subhourly frequency. Eight calibration experiments are carried out, differing in model adequacy, time resolution and discrepancy definition. Results show the inability of MSE to assess model quality when applied to high-frequency measurements. In comparison, it can be argued that the alternative metric based on DTW better reflects model quality.