Introduction

Building energy consumption currently account for around 40% of total energy consumption and 70% of energy demand in the U.S.. By creating tool to help building managers to make decision on increasing building efficiency, it will be one of the easiest ways reduce environmental impact.

Energy benchmarking is a non-labor intensive process on creating a relative energy efficiency metrics by comparing similar buildings based on the characteristics, such as weather, gross area, location. The result can later be translate into an actionable insight for the building manager, to have a better understanding on the saving potential and drive cost effective change.

This project will be applying quantile regression to benchmark the total building consumption for universities in the United States. The methodology is based on research paper by Jonathan Roth and Ram Rajagopal from Stanford University, “Benchmarking building energy efficiency using quantile regression”. Although their original application is on individual buildings, a similar approach can be apply to university campus with slight adjustments. Roths address issues current benchmarking methods by creating the following practices:
1. Cumulative Distribution Plot that showing the individualized information, such as theoretical maximum performance.
2. Reduces outlier effect. Assuming there has no error in data, the outlier can be representing the very efficient building, therefore by applying the conventional way, which is ordinary lease square (OLS) regression, the end result will be either underestimated or overestimated.

This document uses Stevens Institute of Technology as an example to provides guidance on interpreting the dashboard. The example contains both baseline and performance year data, while the dashboard only contains performance year data at the current stage.

Data

The dataset is extracted from The Association for the Advancement of Sustainability in Higher Education (AASHE), which contains sustainability related data of over 200 universities across United States and Canada. The features of the dataset are selected based on the result of research by Roth, 2018. If time is allowed. this project will perform feature selection with additional variables.

Extraction of Dataset

School Consumption Employees Student Health Cool Area Heat Lab
Agnes Scott College 63336.1 418 917.000 5601 1550 1022795 3148 25492.0
American University 316968.9 3338 12504.000 4815 1908 4553432 3311 107561.6
Antioch College 25589.0 154 133.000 0 900 388246 4858 20000.0
Arizona State University 1916420.6 10460 87955.000 30129 5071 23316686 829 1622712.0
Auburn University 1216793.0 5873 24849.000 163336 2331 11846397 1809 478837.0
Austin College 105806.7 338 1232.143 3900 2677 1011199 2344 29925.0

Data Description

Data Name Brieft Description Unit
Consumption Total Building Energy Consumption MMBtu
Employees Number of Employees (Faculty + Staffs) N/A
Student Number of Full time Student N/A
Health Total Area of Healthcare sqft
Cool Cooling Degree Day (65°F Above) Degree-Days (°F)
Area Total Gross Building Area sqft
Heat Heating Degree Day (65°F Below) Degree-Days (°F)
Lab Total Area of Labortory sqft

Correlation Matrix

Operational characteristic of Stevens Institute of Technology

Year Number of Employees Number of Student Area of Healthcare (sqft) Coolday Degree Total Gross Area (sqft) Heatday Lab Area (sqft)
2015 1255 5129 0 1348 1480090 5172 131359
2019 863 6929 0 1454 1480090 4773 131359

Methodology

There are several limitations and problems with the existing energy efficiency benchmark products and studies, including partial set of explanatory variable, low resolution data and using solely energy use intensity as a primary metrics. The following section will further discuss how the propose methodology solve the limitation to the existing problem:

  1. Using energy use intensity as primary metrics to rank university can be implement with ease and no further analysis is necessary. However, this method neglect other characteristic of the university that contribute heavily to the energy usage variation between university, such as weather pattern, existence of gym or lab.
  2. Ordinary lease square (OLS) is the most commonly use regression method, Energy Star is using the technique to benchmarking the building energy efficiency. The problem with OLS is the high sensitivity to outlier, and can skew the regression line towards them. This method also does not provide Individualized recommendation, such as theoretical maximum performance of the university.

Proposed Benchmarking Model

The following section will briefly explain the proposed benchmarking model with the help of flowchart (Roth, 2018), and follow by step by step text description.

  1. Quantile Regression: Quantile Regression is chosen because this method is useful when extreme is important, while in this case is to identify the best and worst energy efficient universities. It provides a complete picture with conditional distribution for each quantile (tau value). Unlike Ordinary Lease Square (OLS), it does not require the data to have normal residual and constant variance.

  2. Individualized cumulative distribution function (CDF): The second step of the benchmarking model is to generate cumulative distribution function for the Stevens Institute of Technology. CDF plot can be interpret as theoretical distribution for each building.
    By applying quantile regression on energy benchmarking, 99 models will be build (tau value from 0.01 to 0.99. In this case, tau value is from 0.05 to 0.95). Those models will later be aggregate conditional distribution across quantile, to form cumulative distribution plot for the target university based on the given input. From the chart, campus manager then able to realized the saving potential, and to set saving target based on the quantile (score) target to be achieve.

  3. Quant Rank: Quantile Ranking can be obtained from the CDF plot, by finding the closest predict energy consumption with the actual energy consumption. The corresponding tau value will be the quantile (score) of the university. For example, when the corresponding tau is 85, then you are at \(85^{th}\) percentile, way less efficient among other university relatively.

  4. Influence Plot: Influence Plot is the plot of coefficient respect with quantile for each features from the model. The influence of each feature on the energy efficiency can be observed.

Result

## Using School as id variables

Implication

Based on the actual energy consumption of Stevens Institute of Technology (SIT), the score can be obtained from the cumulative distribution plot. Using the information from year 2015 and year 2018, the scores are 71 and 92 respectively. It is alarming signal to the campus manager that the campus is becoming less energy efficient, and consuming more energy than their comparable peers that sharing the similar characteristics. The significant decrease in energy efficiency may due to the lack of change total gross area of building data from 2015 to 2018, therefore buildings that is built in between year 2015 and year 2018, such as Babbio Garage and new construction site are not include in the data analysis process.

The density plot shows the density of quantile point respect with the change in normalized energy usage. The higher the density at a specific level, it is more easier for the campus manager to alter the energy efficiency score. Comparing with year 2015 and year 2018, it is more difficult in year 2018 to increase energy efficiency. The tipping point for both years occur on the left end, and for year 2018 is further left on the chart, which means it is more difficult to reach the tipping point from the current level.

Reference

Jonathan Roth, Ram Rajagopal, Benchmarking building energy efficiency using quantile regression, Energy, Volume 152, 2018, Pages 866-876, ISSN 0360-5442, https://doi.org/10.1016/j.energy.2018.02.108. (http://www.sciencedirect.com/science/article/pii/S0360544218303360)