A novel Energy-Flow based Ensemble Calibration model for rapid and accurate energy-economic performance prediction of deep energy retrofit in single-family houses

Abstract

In Europe, 68% of residential building stocks are single-family houses, of which roughly 54% are constructed before the 1970 s with poor energy efficiency. However, few studies have addressed the issue of ‘Rapid and accurate energy-economic performance prediction of deep energy retrofit in single-family houses’ over the past decades. Recently, this issue has been discussed in limited research via Black- or White-box based building energy models to predict post-retrofit energy-economic performance. Applying the Black-box models, low accuracy occurs due to inadequate data, neglect of coupling effects in deep energy retrofit combinations and lack of complex heat transfer phenomena in single-family houses. Conversely, the White-box models with time-consuming and numerous dynamic simulations are very hard to be implemented at a building cluster or urban scale.This paper proposes a novel Energy-Flow based Ensemble Calibration model to tackle this issue, with a maximum discrepancy of 6% compared to other building performance simulation results and 0.72 s computing time on a single combination. This model consists of three layers 1) input neurons, 2) complex heat and electrical transfer models, and 3) output neurons. By replacing the layer of hidden neurons with complex heat and electrical transfer models in conventional Black-box models, the coupling effects in deep energy retrofit combinations and complex heat transfer phenomena are thoroughly considered in this research. Besides, the computational efforts are substantially reduced with only a few building performance simulations required for baseline building and identified individual retrofit measures. A semi-detached 1960 s Italian single-family house is selected as a case study to validate the calculation results from this model against real energy consumption of baseline and post-retrofit scenario, with a discrepancy of −0.66% and 1.03%, respectively.