Lawrence Berkeley National Laboratory

Utility programs have successfully delivered energy efficiency for decades. Today, increasing emphasis is being placed on demand response (DR) programs that incentivize customers to reduce, or “shed” electric load during grid peak periods. The most common methods used to predict building peaks and quantify DR load reductions rely on simple averaging algorithms using hourly load and temperature data from the days preceding the DR event. In contrast, regression-based algorithms have been used for decades to quantify annual energy efficiency savings. The availability of smart meter data has enabled application of hourly regressions for more accurate energy savings estimation, often referred to as “advanced measurement and verification (M&V).” This project explored whether advanced M&V regression approaches offer improvements over simpler averaging approaches for peak load prediction in commercial buildings. We present evaluation results for eight algorithms (based on three baseline modeling approaches). The findings show that all algorithms underpredicted consumption across 453 meters and over 1,100 peak load days. Median bias values varied between 4.5 and 18.7 percent, indicating that the methods evaluated would tend to understate achieved load reductions in DR applications for these buildings. The regression methods did not offer a notable advantage over the commonly used averaging methods.