Commercial and residential buildings account for approximately 70% of the electricity consumption 40% of the primary energy consumption in the U.S.  Managing these demands while minimizing cost and greenhouse emissions requires efficient integration of distributed generation (DG) and storage, development of suitable demand response strategies, and planning and management of electricity markets at a regional level. All these measures require electricity and heat demand predictions over a medium-to-long term time horizon (i.e. \> 1 week). Such longer-term (i.e. \> 1 week) predictions of electricity at a high-resolution can also be used for economic dispatch in transmission networks; whereas predictions of heat demand can be used to estimate emissions from buildings that can be potentially be coupled with climate models. Physics-based energy simulation packages are often used to estimate these energy demands; however, these computational tools often require detailed information on building construction and schedules that are often not available in practice.  
  
Therefore, there is a need to develop machine learning (ML) models for making medium-to-longer term forecasts of electricity and heat consumption. The presented work develops a novel deep recurrent neural network (RNN) model to make longer term electricity and heat predictions at one-hour resolution by leveraging the sequential behavior of energy consumption. The RNN models the sequential behavior by using an encoder-decoder architecture with long short-term memory (LSTM) units., which allows the RNN to account for long-term temporal dependencies in the electricity profile, and to surrogate for the temporal effects of building schedules. The accuracy of the RNN model was first evaluated using multiple electricity profiles obtained from the Public Safety Building in Salt Lake City, UT, USA with each profile corresponding to a separate end-use. The RNN performed comparatively better than a multi-layered perceptron (MLP) neural network – particularly in the case of HVAC profiles. The RNN model was subsequently evaluated on heat demand profiles in multiple campus buildings at the University of Utah, and in general, the RNN model performed comparatively better than the benchmark MLP model.

The formulation of the above-mentioned RNN model led to two additional applications pertinent to energy forecasting. The first application of the RNN was to solve the missing-value problem, i.e. how to impute missing values when large sequences of training data are missing or corrupt, a common occurrence in time-series building energy data. We proposed a simple imputation scheme using the RNN model to address this problem. The second application of the RNN model was to investigate how the longer-term predictions at one-hour resolution can be used to make predictions at a sub-hourly interval (i.e. 10-minute resolution). In this work, we present and evaluate multiple strategies for using the RNN model, predicting at a one-hour resolution, to inform a second RNN model making predictions at a 10-minute resolution.

Methods and results from this study can provide useful guidance on subsequent research pertaining to data-driven energy forecasting and optimization. Future work can potentially focus on: (i) developing methods to compute uncertainties associated with RNN predictions; (ii) investigating methods with which ML models can be integrated with different modules in energy simulation packages.     
  
 


Predicting Electricity and Heat Demand in Buildings Over a Longer Time Horizon Using a Sequence-to-Sequence Recurrent Neural Network Model

Division / Sponsor:
Advanced Energy Systems Division (AESD) & Solar Energy Division (SED)

Paper Proceedings: https://asmedigitalcollection.asme.org/ES/ES2020/volume/83631

ES2020 - Virtual Conference 

Supporting the need among researchers worldwide for additional online and multimedia content, ASME now offers the ASME Conference Video Collection. This platform will collect and provide long-term access to conference presentation videos, encompassing all technology areas within ASME's technical conference and events program. About 30,000-40,000 engineers across all mechanical engineering and related disciplines participate in ASME conferences each year. This includes authors of approximately 6,000 to 8,000 technical papers published by ASME, the majority of whom also submit video presentations for our virtual conferences. The video collection will be available as an add-on to existing product packages or as a separate subscription.