Gaussian-Process-Based Emulators for Building Performance Simulation

In this paper we present a novel emulator of a building simulator for the simulation-assisted design of high performance buildings. Our emulator is based on Gaussian-Process (GP) regression models. Such non-linear models are better suited than linear models as emulators because the simulator itself is a collection of non-linear models based on differential equations. We show that our proposed emulator is about 3 times more accurate than linear models in predicting the output of the simulator, achieving an average error of around 10-25 kWh/m2 for prediction of energy outputs that are in the range of 10-800 kWh/m2, compared to an error of around 50-100 kWh/m2 obtained by using linear models. Our emulators also heavily reduce the computational burden for building designers who rely on simulators. For example, the emulator can first be trained with observations from the simulator using a wide variety of building designs and weather data. This pre-trained model can then be used by building designers for exploration of new designs by predicting the performance of new buildings very quickly (in just a few milliseconds). We expect our approach to be particularly useful for Uncertainty Analysis (UA), Sensitivity Analysis (SA), robust design, and optimisation.

Generation of Weather Files Using Resampling Techniques: An Exploratory Study

Marilyne Andersen, Parag Rastogi

Simulating a building to predict its performance over the course of a full year requires an accurate representation of the stable and representative weather patterns of a location, i.e. a weather file. While weather file providers give due consideration to the stochastic nature of weather data, simulation is currently deterministic in the sense that using one weather file always generates one performance outcome (for a given set of building parameters). Using a single time series or aggregated number makes further analysis and decision-making simpler, but this overstates the certainty of the result of a simulation. In this paper, we investigate the advantages and disadvantages of incorporating resampling in the overall simulation workflow by comparing commonly used weather files with synthetic files created by resampling the temperature time series from the same weather files. While previous studies have quantified uncertainty in building simulation by looking at the calculation itself, this paper proposes a way of generating multiple synthetic weather files to obtain better estimates of expected performance. As case studies, we examined the performance of the ‘original’ and synthetic files for each of a sample of world climates.

International Building Performance Simulation Association2013

On the sensitivity of buildings to climate

Parag Rastogi

Building simulation requires a large number of uncertain inputs and parameters. These include quantities that may be known with reasonable confidence, like the thermal properties of materials and building dimensions, but also inputs whose correct values cannot be known with absolute certainty, notably weather and occupancy. A simulation run is not, strictly, a prediction. Since the parameters and calculations are approximations of real-world phenomena and materials, the exercise is essentially uncertain. Regardless of whether simulation is interpreted as a prediction or an approximation indicative of average behaviour, including explicit bounds of uncertainty is more informative for a decision-maker than a single point estimate. This thesis presents results for two related but independent proposals for sensitivity and uncertainty analyses in building simulation, particularly to weather. The first is a novel, generalisable procedure for generating synthetic weather data to carry out a Monte Carlo experiment with a building simulation model. The second is a technique for training emulators or response surfaces to rapidly obtain estimates of performance outputs from simulation models, using Gaussian Process regression on small training data sets. The two parts, together and separately, enable the quantification of the lack of knowledge about an input, and the impact of this uncertainty on the final results. The synthetic weather time series developed are an ensemble of realistic hourly data whose mean statistical characteristics are close to the typical year used to generate them. The procedures developed are generalisable with minimal expert input. We avoid presenting a unified model for all climates, leaving some tuning parameters like the extent of correlation, and the unknown coefficients of stationary time series models, to be calculated empirically (based on the typical file of a given climate). The emulators are created using regression, comparing the performance of classical parametric regression with a non-linear technique based on Gaussian random processes. Our proposal trains reliable models on small samples, reducing the computational burden, and gives an explicit estimate of the uncertainty for a prediction, since the response at any sampled point is modelled as a Normally-distributed random process. Once again, we avoid a unified emulator or regression model because the response from one building (defined by its geometry and usage in this case) is not necessarily an appropriate description of the response of another. This work is a step towards practical tools for the use of building simulation in a stochastic paradigm. Both elements of the thesis contribute toward explicitly estimating the uncertainty in the results of building simulation, using empirical or data-driven techniques. The types of the time series and emulator models are general enough to work on any climate or building, with parameters obtained from the simulated/typical sample at hand, but the importance of different aspects and the nature of a buildingâs response are determined uniquely (i.e., parameter values). The work is easily extensible to the analysis of the sensitivity of a building, or groups of buildings, to any inputs. The concepts proposed in this thesis may also be used for stochastic optimisation and models to predict performance metrics other than the annual sum of energy.

EPFL2016

On the sensitivity of buildings to climate

Parag Rastogi

EPFL2016