DemandData
A Gradient Boost model for building-level electricity load forecasting
A multiscale approach to electricity demand estimation at half hourly resolution. Statistical information relating to the building stock and demographics of a target geographic region of interest (ranging from a single home, commercial or industrial building) up to any larger and arbitrary Electricity Supply Area (ESA) could be used to estimate electricity demand. A model was designed which would train using aggregate demand time series data for e.g. Grid Supply Points (GSPs) and corresponding aggregated building stock and demographics statistics in order to predict the disaggregated demand profile for e.g. primary substations.
Geomni and Experian data were to be joined at premise level since in the UK each premise has an individual Unique Property Reference Number (UPRN). The combined dataset was then aggregated to GSP, Bulk Supply Point (BSP) and Primary Substation before being vectorised ready for application in machine learning.
Electricity demand metering data was also to be assembled and assigned to supply regions within which the statistical distribution of the building stock and demographics attributes (features) could be computed.
The prototype was to demonstrate the fitting and application of a model that could estimate half hourly electricity demand time series for arbitrary regions of interest based on the statistical features within the region and a higher level demand time series associated with a larger region of interest. The larger regions could be those associated with a particular GSP, BSP or primary substation - depending on the metering data available associated with the smaller region of interest.