Let me start by welcoming the reader to Rebate Bus — we’re working to make energy efficiency incentives more efficient by cataloging them alongside lighting product data to deliver an integrated experience to commercial property owners. You can read more about the company at www.RebateBus.com
One issue that we’ve run across recently relates to the idea of equivalent wattage. For a given light bulb or fixture, equivalent wattage is defined as the likely wattage of a product using an older technology, such as incandescent or halogen, which provides a similar amount and type of light for a similar application. Equivalent wattage with respect to a given market is thus heavily dependent on exactly what is installed currently in that market. However, for utility incentive purposes, it is important to have a general idea of what the energy savings impact will be when a product is sold. Thus we at Rebate Bus, having a mission to make such incentives more efficient, have an interest in seeing the equivalent wattage information accessible to utilities be as accurate as possible and in estimating it given the details of a product. I set out yesterday to build a predictive model using my knowledge of statistics and machine learning.
Current information in the market on equivalency is limited — ENERGY STAR, Design Lights Consortium and Lighting Facts are the main sources of data on products, and of the three only ENERGY STAR provides equivalent wattage. We began our development of a predictive model by testing the performance of a regression using polynomial features on the ENERGY STAR data. For this regression, we split the 11819 products listed into two sets: a training set composed of 5000 products and an evaluation set composed of the remaining products. We used 3 features in the predictive model: luminosity, wattage, and wattage squared.
Fig 1: Visualizing the results of an equivalent wattage regression
This model performs moderately well — the mean experimental error is 10.64 watts. Such an error is not the end of the world, but it is problematic. How problematic? Well, if we assume that the light bulb in question operates for 12 hours a day, a 10.64 watt error will throw off our estimate of kilowatt-hours saved by 10.64 * 12 * 365 / 1000 = 46.6 kWh. With typical incentive levels around 12 cents per kilowatt-hour saved, that means that we’d estimate the utility would pay $5.60 more than they actually would — a figure that can really add up for a big project. Another, more problematic issue with the model is that it is tuned for ENERGY STAR products. These products use relatively low wattage and are limited primarily to light bulbs whereas Rebate Bus handles all fixture types. In the next installment, we’ll aggregate some data and take another look at the regression in order to try and improve these results.
Email me at firstname.lastname@example.org if you’re interested in seeing more about what we are doing here at the Rebate Bus.