Full Library

Weather data is a crucial input for myriad applications in the built environment, including building energy modeling and daylight analysis. Building science practitioners and researchers have been able to select from a variety of weather files, such as Weather Year for Energy Calculation 2 (WYEC2) and the Typical Meteorological Year (TMY). However, commonly used weather files are typically synthesized to represent trends over a relatively longer periods of time, and are often unable to accurately depict climatic conditions that result from local contexts, such as the heat island effect, wind flow, even local temperature and relative humidity. This results in discrepancies in building performance simulations. This study proposes a methodology using recurrent neural networks to generate synthetic localized weather data that are significantly more accurate and representative of local conditions than standard weather files. The predictions were validated against actual on-site measurements, and achieved a low mean square error of 2.96 and over 185% improvement in validation accuracy. Overall, the performance of selected models has shown over 100% improvements in test accuracy compared with standard weather files and weather station data at the nearest airport. The proposed methodology can be used to morph generic weather files to accurately represent localized conditions, or generate localized data for a longer time span with only a subset of data available/collected. This is useful for downstream built environment applications, especially building energy modeling, since representative weather data capturing trends of temperature and other variables will result in enhanced accuracies of the building energy models. The method can also be used in urban analysis pipelines to enhance resilience against climate change.

openalex: W3120887079