Neural network-based nonintrusive calibration for an unobserved model in digital twin-enabled building operations

Abstract

The digital twin (DT) model is the key to enabling DT-enabled applications and services by mathematically expressing the physics, dynamics, and indoor environment of a target building. However, the DT model may provide substandard performance owing to various operational uncertainties; as such, in-situ calibration is essential in building operations. This study proposed a neural network-based nonintrusive calibration (NNNC) method for a DT prediction model. A field application study was conducted on a prediction model developed for a district heating substation serving real residential buildings in Korea. The case study demonstrated (1) the operational uncertainty of the prediction model and (2) the dependence of the calibration accuracies on different indirect calibration approaches. The NNNC method showed a better calibration accuracy for the target prediction model than existing Bayesian inference-based nonintrusive calibration. Specifically, NNNC, using a correction function with a multi-layer perceptron, reduced the prediction model errors by 38.6% and 34.7% for bias and random errors, respectively. The results demonstrate the effectiveness of NNNC in DT-enabled building operations. Therefore, the NNNC ensured the applicability and reliability of the prediction model.