EMBEDDING STOCHASTICITY IN BUILDING SIMULATION THROUGH SYNTHETIC WEATHER FILES Parag Rastogi 1 and Marilyne Andersen 1 1 Interdisciplinary Laboratory of Performance-Integrated Design (LIPID), Ecole Polytechnique F´ ed´ erale de Lausanne (EPFL), Lausanne, Switzerland ABSTRACT This paper presents an attempt to create synthetic weather data for stochastic building simulation. The synthetic data are created entirely from the freely available Typical Meteorological Year (TMY) weather files using time series models and resampling. The generated data turn out to be representative of recor- ded data for our case study without any prior ‘know- ledge’ of the long term distributions of meteorological parameters. The current model does not address spells above or below some temperature of interest (e.g. heat waves), and the authors are working to incorporate that in future work. Another avenue for further exploration is modifying the mean to incorporate the results of Re- gional Climate Models for future conditions. Correla- tion of the synthetic data with synthetic solar radiation and humidity has been verified and the authors’ work with this ensemble of weather time series of interest will be presented in future publications. INTRODUCTION Using TMY files in building simulation for ‘what-if’ analyses of designs tells us only about the response of a building or design strategy to typical climatic condi- tions, for the period of record of the file. However, sev- eral studies have pointed out the sensitivity of simula- tion output to weather data, including Bhandari et al. (2012), Chinazzo (2014), Crawley and Huang (1997) and Hong et al. (2013). In other words, simulating with typical weather gives no information about the sensitivity of a building or design strategies to vari- ations in the climate itself. The difficulty of fully characterising a system (for sensitivity or uncertainty analyses) that depends on the climate is that we cannot fully characterise the cli- mate itself, especially future climate. This epistemic uncertainty has led some to propose that a ‘range’ of possible performance outcomes, i.e., the results from simulation runs with different weather inputs, better characterise the range of performance that a building will inevitably give (e.g., Chinazzo et al. 2015). If one does not know exactly what (weather) inputs one’s (building) system will experience, then one is better off knowing the effect on it of a range of possible in- puts. A given weather file is, after all, a representation of one scenario out of an immense number of possib- ilities. Therefore, by using only one weather file, we are restricting ourselves unnecessarily to one “experi- mental result”. If a building never experiences a nar- row set of weather conditions exactly, i.e., the ones contained in a typical weather file, then the quality or ‘averageness’ (or ability to represent best the most typ- ical weather) of said weather file is irrelevant. The aforementioned studies usually propose using measured weather data from the vicinity of the build- ing to best characterise the climate. While long simu- lation with measured data is, intuitively, a better depic- tion of climatic variability than a single ‘typical’ time series, it can not guarantee coverage of future condi- tions or extremes. While the historical distribution of a weather parameter can be known from recorded data (usually at a resolution of one day or month), using historical records to study the future represents an as- sumption about the future: that of a stable climate. The IPCC’s latest Synthesis Report (IPCC Core Writing Team 2014) is unambiguous in its assertion that the climate is changing, though it is not knowable which of its future scenarios best represents how the global climate will evolve. Assuming one has access to long-term hourly data from a weather station that is sufficiently close to the area of interest, in addition to a TMY-type file, one can know how a building would have behaved. However, one has no tools for assessing any arbitrary weather conditions. One might have a reasonable idea of the expected range of average temperature rise in a cli- matic region, thanks to the IPCC’s publicly available models, but one does not know the possible implica- tions of this at an hourly resolution for a given weather station. Our approach seeks to address this incertitude by proposing a ‘what-if’ analysis of a building to vari- ations in the climate, without seeking to forecast the ‘true’ future climate. The first step in the development of this climate sens- itivity analysis procedure is introduced in this paper: stochastically-generated synthetic input data. We ex- tract the essential characteristics of a climate (e.g., autocorrelation, means, etc.) and build any number of synthetic files by modelling the structures and reshuff- ling the apparently random components of the time series. This is possible because of the idea, developed by several authors (including, Boland 1995; Boland et al. 2013; Hansen and Driscoll 1977; Magnano et al. 2008; Scartezzini et al. 1990), that weather time series