Operating at the leading edge of building energy performance analysis, Buildings Alive commits a significant amount of resources to research and development activity. Several of Buildings Alive’s staff have had research papers accepted to the prestigious American Council for an Energy Efficient Economy (ACEEE) Summer Study in August 2018.
The accepted papers and their abstracts are:
Predicting peak energy demand in commercial buildings under extreme conditions: by how much can we improve accuracy?
Hao Huang and Craig Roussac
Energy demand predictive models are an important tool for both automated and behavior-based demand response (DR) in commercial buildings. Machine learning methods, like support vector regression (SVR) and ensemble tree methods, are effective at solving non-linear stochastic modelling problems, and have therefore been widely used for energy prediction. However, machine learning models trained using historical weather and historical demand data have the weaknesses of not generalizing well when historical data sets are small. It is even more difficult to make an accurate prediction when the forecast inputs are beyond the range of training data. This makes it challenging for machine learning models to generate accurate peak demand predictions under extreme weather conditions. This paper presents a novel model structure to predict peak demand more accurately under extreme conditions. Five machine learning methods, both linear and nonlinear, are investigated and compared to obtain the most suitable model for extreme day correction. The proposed modelling approach has been automated at a computer server for application in a behaviour-based demand response program across ~80 different buildings in Australia. The real-time modelling results from 2016–2017 summer have shown an overall ~3% to 6% of improvement in mean absolute percentage error (MAPE) compared to a highly accurate benchmark model. This has resulted in more accurate warning delivery and a higher demand response rate. The aggregated results in 2018 shows the model can accurately detect 90% of energy demand peaks up to five days ahead
For Building Operators, What Difference Does a Target Make?
Craig Roussac and Hao Huang
There is increasing recognition, both in the literature and from policy-makers, of the influence that operations personnel have over the energy efficiency of large complex buildings and the potential to save energy by providing them with clear and timely feedback. An implicit but poorly-founded assumption built into most studies of energy efficiency feedback is that participants have a goal to save energy. Often this is not true and even among those who seek to achieve savings there is much variability in how competing goals are prioritized. In this study, operators of 140 large office buildings received daily feedback informing them about their buildings’ energy efficiency compared to an “expected” (baseline) level determined by reference to a statistical model that normalized for weather and other factors outside of their control. Variances between expected and actual performance gave them signals about the impact of their actions on energy use. After a period, the focus of the daily messages was switched for 60 of the operators from a baseline profile to the ‘best’ normalized performance observed previously. The introduction of this daily “target” profile was directly associated with additional savings. It was noted that engagement with the daily feedback increased, even though participants had no role in target setting. This study has implications for the design of information feedback methods for building operators and suggests that by giving them an awareness of the divergence between their building’s performance and its efficiency improvement potential, a positive spiral of increasing energy efficiency, expertise and engagement can emerge.
Forewarned Is Forearmed: Reducing Peak Demand in Commercial Buildings with Behavior Science.
Craig Roussac and Hao Huang
As peak demand pricing becomes commonplace, there is an opportunity for the engineering operations managers of commercial buildings (“operators”) to reduce electricity demand charges by shifting peak load. Energy prediction based on historical data can provide useful information to building operators by allowing them to foresee, and therefore manage, their peak electricity use. Although energy modelling technology has been broadly discussed, no study has been conducted to examine the potential to utilize forecast information to change operators’ behavior in managing peak demand in buildings. This paper presents findings from a field study of operators’ responsiveness to peak demand warnings, and the impacts of their behaviors on their buildings’ electricity demand. The study was performed at 71 buildings in Australia during the 2017/18 summer (23 in Melbourne and 48 in Sydney). Each building’s energy profile was forecast five days in advance and a warning message was issued if the predicted demand exceeded a pre-defined threshold. A summary comparing the divergence between expected and actual demand was sent the day following the predicted peak demand day. Continuous engagement among engineers, tenants and building operators was also conducted. By comparing the buildings’ 2017/18 maximum demand with 2016/17, and taking into account ambient weather conditions, we found that the operators produced significant peak demand reductions in Sydney (2.98MW) which could be explained by their direct actions. In Melbourne, peak demand increased (by 1.16MW) on account of a prolonged heatwave despite evidence of the operators’ active engagement. We estimate the effect of the intervention was to reduce peak demand on average by 3W/m2 (0.28 W/ft2)
Unsupervised fault detection and diagnostics: detecting unusual behaviour in buildings using BMS data
Cameron Roach
Building Management Systems (BMSs) are used to control HVAC equipment, lighting and other devices in commercial buildings. These systems can generate significant volumes of data with a single building typically containing tens of thousands of sensors. Due to the large volumes of data it is difficult for facility managers to quickly assess if a building is performing as expected or if faulty sensors are present. Furthermore, the data that is collected can often be difficult to deal with. Irregular time intervals, missing values, outliers and inconsistent sensor labelling each add complexity and introduce new problems that need to be addressed if accurate fault detection is to be carried out. This paper explores using an unsupervised machine learning methodology to allow end users to quickly assess if a BMS is behaving as expected. To deal with the inherent complexity of the time series data and metadata we engineer useful data features to help improve our analysis. We test several dimensionality reduction techniques that allow us to visualise data easily. We also assess different clustering approaches which can be helpful in detecting if any sensors are behaving erratically or if any sensors have been incorrectly labelled. Applications of this methodology include fault detection and improving our understanding of BMS data. We also examine its performance across multiple buildings and systems and discuss extensions such as the possibility of comparing behaviour across multiple buildings.