A Conditional Generative adversarial Network for energy use in multiple buildings using scarce data

Building consumption data is integral to numerous applications including retrofit analysis,Smart Grid integration and optimization,and load forecasting.Still,due to technical limitations,privacy concerns and the proprietary nature of the industry,usable data is often unavailable for research and development.Generative adversarial networks(GANs)-which generate synthetic instances that resemble those from an original training dataset-have been proposed to help address this issue.Previous studies use GANs to generate building sequence data,but the models are not typically designed for time series problems,they often require relatively large amounts of input data(at least 20,000 sequences)and it is unclear whether they correctly capture the temporal behaviour of the buildings.In this work we implement a conditional temporal GAN that addresses these issues,and we show that it exhibits state-of-the-art performance on small datasets.22 different experiments that vary according to their data inputs are benchmarked using Jensen-Shannon divergence(JSD)and predictive forecasting validation error.Of these,the best performing is also evaluated using a curated set of metrics that extends those of previous work to include PCA,deep-learning based forecasting and measurements of trend and seasonality.Two case studies are included:one for residential and one for commercial buildings.The model achieves a JSD of 0.012 on the former data and 0.037 on the latter,using only 396 and 156 original load sequences,respectively.

Using Bayesian deep learning approaches for uncertainty-aware building energy surrogate models

Fast machine learning-based surrogate models are trained to emulate slow,high-fidelity engineering simulation models to accelerate engineering design tasks.This introduces uncertainty as the surrogate is only an approxi-mation of the original model.Bayesian methods can quantify that uncertainty,and deep learning models exist that follow the Bayesian paradigm.These models,namely Bayesian neural networks and Gaussian process models,enable us to give predic-tions together with an estimate of the model’s uncertainty.As a result we can derive uncertainty-aware surrogate models that can automatically identify unseen design samples that may cause large emulation errors.For these samples the high-fidelity model can be queried instead.This paper outlines how the Bayesian paradigm allows us to hybridize fast but approximate and slow but accurate models.In this paper,we train two types of Bayesian models,dropout neural networks and stochastic variational Gaussian Process models,to emulate a complex high dimensional building energy performance simulation problem.The surrogate model processes 35 building design parameters(inputs)to estimate 12 annual building energy perfor-mance metrics(outputs).We benchmark both approaches,prove their accuracy to be competitive,and show that errors can be reduced by up to 30%when the 10%of samples with the highest uncertainty are transferred to the high-fidelity model.