Automated residential energy audits and savings measurements using a smart wifi thermostat enabled data mining approach

The building sector has been identified as one of the biggest contributions to electricity and natural gas consumption in the U.S. These findings have necessitated the need for the development of energy saving initiatives in the sector, which will aid in reducing greenhouse gas emission needed to reduce the risk of climate change. However, despite several efforts by state agencies, such as the implementation of Property Assessed Clean Energy (PACE) and On-Bill Repayment or On-Bill Financing of energy efficiency investments, there are significant challenges to achieving energy efficiency in the building sector. Fundamentally the question is “How do we find the most cost effective energy efficiency measures present in the world?” Conventional energy audits, the typical way to discern, struggle from high cost, inconsistency in audit recommendations, and a lack of people trained to deliver. Thus, the approach just is not capable of “at-scale” identification of the measures to address first, then second, and so on. Additionally, it is essential that the savings from any investment and/or even behavioral changes be capable of being measured with accuracy in order to improve the ability to find the most effective energy reduction measures existing in the broader building sector and in order to communicate the relative economic benefits from upgrades to building owners. At this time, unless there are short-interval energy meters in buildings, the ability to measure savings with accuracy is just not there. As a solution, this dissertation investigates utilizing smart Wi-Fi thermostats data to conduct visual energy audits and predict energy savings with improved accuracy from any energy systems upgrade and any behavioral modification. The study leverages data from 101 residences owned by the University of Dayton. In 2015 prior University of Dayton researchers completed energy audits of these; documenting the geometric and energy characteristics and occupancy, as well as documenting any unique energy consuming device such as washers/dyers/dishwashers in the residence. These houses provided a diversity of size, age, insulation, and energy effectiveness. Additionally, historical energy consumption data, as well as smart WiFi thermostat data with corresponding weather data, were collected for these houses. The archived thermostat measured temperature data was used to develop unique power spectrums for the measured interior temperature for each residence. The binned power spectral density is shown to be an effective signature of the energy effectiveness of the various energy characteristics associated with a residence. Moreover, the outdoor temperature for each meter period was binned into histogram groupings. This research utilizes an AutoML H2O package to determine the best machine learning algorithm for predicting both the energy characteristics and energy consumption, as well as complete the tuning needed to determine the best model hyperparameters. Machine learning models were trained to predict attic and wall R-Values, furnace efficiency, and air conditioning seasonal energy efficiency ratio (SEER) using smart WiFi thermostat measured temperature data in the form of a power spectrum, corresponding historical weather and energy consumption data, building geometry characteristics, and occupancy data. The models validation coefficient of performance (R2 values) were respectively 0.9408, 0.9421, 0.9536, and 0.9053 for predicting attic and wall R-Values, furnace efficiency, and AC SEER. This research helped lift up the possibility of conducting low-cost, large-scale, data-based energy auditing of residences that rely only on data that could easily be collected for any residence. Similarly, a power spectrum derived from the measured thermostat indoor temperature is combined with outdoor temperature data and known residential geometrical and energy characteristics in order to train a singular machine learning model capable of predicting energy consumption in any residence. The best model obtained had a percentage mean absolute error (MAE) of 8.6% for predicting monthly gas consumption. This result indicates that the best model is effective to estimate energy savings from upgrades in residential buildings. Specifically, when it is applied to real residences in which attic insulation upgraded, the energy savings estimation uncertainty was less than 7%. This is a significant improvement over the ASHRAE recommended guidelines for estimating building energy consumptions and savings, which has been termed capable, at best, of resolving savings only greater than 10% of total consumption, and, in many cases, unable to resolve any savings at all