| Description |
The building sector currently contributes to approximately 73\% of the electricity consumption and 39\% of primary energy consumption the U.S. In order to meet these energy needs while minimizing greenhouse emissions, accurate predictions of building energy consumption over a medium-to long-term time horizon (i.e. ≥ 1 week) is needed. These medium-to long-term forecasts can be used to effectively integrate distributed generation and storage systems, as well as incorporate demand response strategies. Physics-based models that use deterministic heat and mass balances, such as EnergyPlus, are traditionally used to calculate electric and heating loads for buildings. However, these models often have a high associated error in practice. Statistical and machine learning (ML) models have been employed in previous literature for short-term forecasts with high accuracy. However, there is a research gap pertaining to predictions of building energy consumption over medium-to long term at one-hour resolution. There is also a need to develop reduced-order models of building-scale thermal storage and propose methods that can utilize these models and predictions of future heating demands to analyze the performance of combined heating and power units. As such, the aims of this research are as follows: Aim I: Investigate the relative accuracies of several machine learning algorithms in predicting fuel consumption in buildings. Aim II: Develop deep recurrent network models for medium-to long-term prediction at high resolution (i.e. one-hour resolution or lower), and for missing-value imputation in energy consumption data. Aim III: Develop a transient heat-transfer model of a building-scale stratified thermal storage tank and investigate how this model can be used for sizing a building-scale thermal storage. Methods and results from this research can provide medium-to long-term forecasts of building energy consumption at a resolution down to one-hour which will allow for improved decisions pertaining to building energy efficiency, more effective smart-grid operation and improved estimation of greenhouse emissions. Results from the study will also provide a definitive framework for future work focusing on developing methods to quantify uncertainties associated with future prediction and integration of machine learning with deterministic building energy simulation models such as EnergyPlus. |
