Data-driven approach to generate test reference year weather files for building energy simulations

Test Reference Year (TRY) weather files are widely used in building energy simulations. The ISO 15927 approach of TRY file generation assigns equal weights to primary climate variables - dry-bulb temperature (T), relative humidity (RH) and solar radiation (GHI)- while considering wind (W) as secondary. However, the suitability of this fixed weightings for diverse climates and building types needs evaluation. This study develops and evaluates data-driven location-specific relatively weighted weather files (RE-TRY) in place of conventional TRYs. The variable categorisation is first derived using the entropy method, and weights are obtained using the XGBoost machine learning algorithm trained on simulated building thermal demands over multi-year observed data. The efficacy of RE-TRY files is then tested by comparing the simulated energy performance of a calibrated residential building against the equivalent multi-year weather data (1990–2016) for three locations in warm climates. RE-TRYs were found to select between 5 and 7 candidate months distinct from conventional TRYs, with variable weights ranging between 0.11 and 0.76. Benchmarked against multi-year data, RE-TRYs yielded improvements of 1.5 %, 7.0 % and 6.0 % across Aw, Cwa and BSh climate types, respectively, compared to TRYs. This novel approach to generating location-specific RE-TRY for a building type contributes to more accurate and performance-based weather files that are valuable for energy-efficient building design. These would benefit buildings aiming for net-zero energy consumption, given that errors in estimating the demand of the scales observed here could be the difference between success and failure.