A data-driven approach for the disaggregation of building-sector heating and cooling loads from hourly utility load data

Electrification of space heating in buildings, currently dominated by on-site fossil fuel use, will be an essential element of decarbonization. Some electrification of heating is already underway, and although state-by-state adoption is highly heterogeneous, the associated impact on the grid is already being felt. The same has been true for air-conditioning except adoption levels are generally higher. As we expect rapid adoption of electrification, a model that can seamlessly disaggregate the existing electric load into that for heating, cooling and non-thermal uses is essential. We develop a model that captures the building thermal response to quantities such as building floor areas that change over years and weather that changes through the day and through the year. Complexity of occupancy, thermostat settings, diversity in building envelopes or technologies deployed are not explicitly represented, but their effects are captured as hourly changes in the response. The model can then be used to estimate the non-thermal dependent loads. Once such a disaggregation is available, it can be used to estimate new load profiles as changes in floor area or electrified loads or weather occurrences. The model is validated against actual hourly utility loads by re-aggregating the simulated hourly loads for a single year for all load zones in NYISO (boundary aligns with New York), ERCOT (covering most of Texas), CAISO (covering most of California) and some other individual balancing authorities of California (BANC, TIDC, IID, LADWP and WALC) with mean absolute percentage errors (MAPEs) across all between 3.0% and 6.0%. The obtained model parameters are further tested by backcasting hourly load for the past 10-year period without degradation in errors, which suggest the model is promising for forecasting in the long-term. While our results bridge the gap between building level energy simulation and building stock energy prediction, all source data for the present study are extracted from open-access datasets. The model is available as an open-source tool that can be easily applied to any spatial resolution at any geographical locations, as long as load profiles and building census data are available.