Summary

Analysis on the performance of a range of baselines used in Great Britain, with a specific focus on accurately remunerating individual households for the flexibility provided during events, similar to those in the ESO’s Demand Flexibility Service.

Key findings

① Baselines are sensitive to a number of key factors

• Historical consumption data. More historical data, to a point, improves accuracy. Using roughly two weeks worth of data generally results in lower errors - more than that can increase errors as data becomes outdated. If averaging over more historic data, more recent data should be upweighted.
• Time of day. Errors overnight, when electricity demand is typically low, are significantly lower than errors in the evening, when flexibility events are currently needed to manage peaks in demand.
• Season. Regardless of algorithm, shoulder months like April and November have higher errors, possibly due to changing patterns in heating and electricity consumption.
• LCT ownership. Households with no LCTs generally have lower errors than those with them. ML algorithms may add value for households with LCTs or some level of automation, which are likely to increase in future. LCT-specific features are worth considering to improve accuracy.
• Period of day or time of year. Models may under predict in key windows, which may depend on household characteristics.

② There is a key trade-off between accuracy and simplicity

• The key above factors should be accounted for, but one baselining methodology for every customer may not be appropriate.
• There is a balance to strike between improving accuracy and adding unnecessary complexity. Rule-based methods are simple and perform reasonably well to baseline household level consumption today.
• More complicated deep learning / gradient boosting methods may be more accurate, but harder to explain to customers, and hard to get flexibility service providers to implement. These may be more useful for internal remuneration.