Better assessment of load reduction by sampling and machine learning

To facilitate flexibility mechanisms for residential customers, Enedis introduces the ‘panel method’, a machine learning algorithm that allows a precise assessment of the quantity of energy consumption avoided thanks to load reduction. Two groups of customers are determined: a panel of those involved in flexibility mechanisms for which assessment of the load reduction is required (interest group), and a panel of other residential customers (mirror group). The model fits, through a Lasso regression, a load made from an aggregate of loads of the mirror panel to one of the interest groups on a learning period (with no load reduction). This estimator is then used to determine what the load of the interest group would have been should no load reduction had happened. This method allows for high accuracy of the estimates, simplifies calculations and data collection (not all individual data is needed thanks to sampling), and observations and assessment of other effects, such as anticipation and post-reduction effects, or postponed consumption. Robustness of methods used for assessment of load reduction is essential for stakeholders to take part in flexibility mechanisms for residential customers.