Abstract
Accurate distribution transformer sizing requires reliable peak load estimation. Traditional methods, such as applying standard coincidence factors to subscribed powers, are often imprecise as they fail to reflect a DSO’s specific customer behaviour. To improve accuracy, the French DSO GreenAlp leveraged its smart metering infrastructure. However, while smart meters can technically provide detailed 10- or 30-minute load curves, French regulations generally restrict DSOs to using only daily consumption data, except for specific projects. To address this constraint, GreenAlp analysed detailed interval data, i.e. 10- or 30-minute load curves, from 10% of its customers over a limited duration only; and relied on power subscriptions and daily metering data for the remaining customers and time periods. To this aim, a supervised learning model was developed in order to predict peak loads based solely on the legally accessible data. Cross-validation ensured the model’s reliability, and a methodology was designed for handling incomplete customer data. The new approach proved significantly more accurate than the method previously used by GreenAlp, while remaining straightforward to implement once the learning model is trained. GreenAlp has integrated this model into its grid modelling software, replacing the previous load estimation method. It is now used for daily operations, providing a practical and scalable solution for optimizing transformer load assessments.