Estimating power demand shaving capacity of buildings on an urban scale using extracted demand response profiles through machine learning models

With the increasing electricity demand driven by population growth and urbanization, the national power grids are exposed to massive pressure, which leads to potential electrical blackouts. Demand response (DR) programs, incentivize end-consumers to reduce their demand during certain periods (i.e., DR events), provide an alternative to the costlier path of constructing more power plants. To alleviate the risks of peak electrical demand that surpasses the supply capacity, it is of great importance to estimate the Power demand Shaving Capacity (PSC) of buildings proactively and accurately. However, the accuracy of PSC estimation at a large scale is held back due to the lack of detailed building/equipment information. This study proposed a machine learning based method to infer the DR performance of data-scarce buildings on an urban scale through leveraging an accurate prediction model developed to estimate the PSC of a smaller cohort of data-rich buildings. Specifically, we first developed a supervised learning model to accurately estimate the PSC of twenty-eight buildings using the state-of-the-art ensemble algorithm (i.e., XGBoost). This estimation model was built using data of detailed building and system information along with more than 200 historical DR events information across three years. Next, we created DR profiles for these data-rich buildings using unsupervised learning methods (e.g., K-means) to find clusters of their DR power consumption (i.e., the percentage of their baseline power consumption that was used during DR events). The results showed that the best PSC estimation model improves the accuracy of DR capacities by 92% as compared to the estimation models used in the current practice. Then, three DR profiles were identified, which indicated that a large building in New York City (NYC) has, on average, the potential of saving 67 kWh during a four-hour DR event. Furthermore, with these DR profiles assigned to more than 9,000 data-scarce buildings with a footprint of more than 50,000 square feet in NYC, we found more than 4.5 million kWh demand shaving capacity in total if these buildings were enrolled in DR.

Automated summary

This study proposes a machine learning method to infer the DR performance of data-scarce buildings by leveraging an accurate prediction model. The results demonstrate increased accuracy in DR capacities and the identification of potential demand shaving capacity in buildings.