Behavioral Research

We examined previous research and mapped the consumer's thought process for how they use their scarce resources.

Click to enlarge

User Experience

Our design went through several iterations and rounds of user testing.

---

Initial Idea

---

Website Application

---

In-Home Wall Device

---

Tablet Application

Close

Data Sources and Tools

We selected Pecan Street's Dataport, the world's largest source of disaggregated customer energy data.

From the more than 1,200 residences in Texas, Colorado, and California, we decided to focus our research and product on Austin, Texas for the years 2013 and 2014. This narrowed our analysis to approximately 150 Austin Homes. We combined household demographic and survey data such as appliance usage information, building type, number of residents, total square footage, and weather data for our analysis. The electricity data was available in one-minute increments with average real power over the interval in kW. The natural gas data was more limited. The gas meter readings were cumulative and taken in cubic feet at irregular intervals collected by ERT devices at the homes. The water readings were cumulative and taken in gallons by ERT devices in the homes.

With regards to carbon footprints, we used the CoolClimate Network's, a division of UC Berkeley's Renewable and Appropriate Energy Laboratory (RAEL), API. Since our dataset is anonymized, we decided to use the downtown zipcode of 78704 for the carbon footprint calculations.

To access the Pecan Street Data please contact Pecan Street directly.

Forecasting Analysis

Autoregressive Integrated Moving Averages (ARIMA) were used to predict future use of resources.

We used ARIMA models to predict energy usage for each between today and the end of the month. This allowed us to provide real-time predictions for the end-of-month bill.

Model performance was very strong. Tested on 3 separate users for days in winter and summer, our model predicted within 3 kWh of the day’s actual usage.

Summed at the month level (for bill prediction), we had an error rate of less than 1%.

We optimized our ARIMA model by grid searching on two primary parameters: the auto-regressive term, and the moving average term.

Providing accurate end-of-month predictions allows the user to feel confident in the number they are working against. They can set goals and see how they performed against the prediction when the end of the month comes around.

For our “don’t water” recommendation, we looked at precipitation predictions for the following 5 days. Weather data was provided along with electricity usage data in the Pecan Street dataset.