Home Contact
Jackson Associates JA Clients Previous Applications Analysis Experience
Summary Standard Databases 15 Min - Other Options RECS Hourly Databases ZIP Database Options
Electric Utilities 2021 Indiana AB Model Smart Grid Impacts
New SEC Scope 3 Financed Emissions Reporting Requirements MAISY Financed Emissions Accounting Data/Support Notes & Whitepapers
Publications Miscellaneous Items
MAISY RECS Hourly Loads & Emissions Databases
 

Question 7. How is Nearest Neighbor Machine Learning Used to Develop MAISY RECS Residential Hourly Load Databases?

Short Answer: MAISY RECS Residential Hourly Load Databases use a proprietary MAISY nearest neighbor machine learning process that matches RECS household characteristics to household characteristics in MAISY Metered Hourly Loads Databases to estimate daily end-use hourly load profiles. MAISY Metered Hourly Loads Databases include more than 230,000 end-use daily load profiles. End-use hourly loads of matching households are applied to RECs records and calibrated to RECS annual end-use kWh use.

Longer Answer: The MAISY nearest neighbor machine learning process matches individual RECS households with households in the MAISY Hourly Load Meter Databases based on similar demographic, dwelling unit, appliance, and fuel use attributes. The MAISY matching process is a non-parametric k-nearest neighbor (KNN) algorithm with regression-based refinements. While the MAISY KNN/regression refinement process is technically classified as an AI process, it differs from popular current AI applications like ChatGPT in that it is completely empirically based with algorithmic refinements based on the accuracy of the KNN matching process. KNN-related applications have been used for decades to maximize the application of information in existing data sets to data sets with less than the full complement of data available in the existing data sets. In the MAISY RECS application, the existing household/appliance characteristics in MAISY Metered Hourly Loads Databases are matched to the same information in the RECS database to impute individual end-use hourly loads missing in the RECS household records. End-use hourly loads are then aggregated to whole building hourly loads and calibrated to annual electricity use.

Metered database records are identified as belonging to the RECS household’s neighborhood with a weighted distance measure that includes various demographic/income/dwelling unit/appliance/fuel characteristics. Hourly loads from a matching household in the Metered Database are applied and calibrated to the RECS end-use energy use for nine non-weather sensitive (NWS) end uses including water heating, refrigerators, freezers, clothes dryers, oven, dish washers, clothes washers, lighting, TVs, and other end uses.

Application of actual metered household NWS hourly electricity use is critical for developing realistic end-use load NWS profiles because these loads vary in a random way across days and weeks. These variations can result in unexpectantly high individual hourly loads compared to estimates based on assumed uniform patterns applied in engineering-based models. This variation plays an important role in solar, battery, smart grid, micro-grid and other product development, application, and analysis.

Related MAISY Topics