Cluster-based and regression-enhanced forecasting of day-ahead prosumer demand schedules in distribution systems

Due to the increasing penetration of power-intensive low-carbon technologies in low-voltage distribution grids, preventive congestion management gains importance. Knowledge of demand schedules is a crucial prerequisite for successful congestion prevention. However, since only a subset of customers will voluntarily provide demand schedules, additional forecasts are necessary. This paper proposes a hybrid methodology to forecast day-ahead demand schedules under partial participation. First, representative schedules are generated using density-based clustering of participating customers to provide a baseline forecast of non-participating customers. Second, if additional measurement data is available, the cluster-based forecast is enhanced using a gradient boosting regressor that utilises additional information about historical demand, weather, and price information. The methodology is evaluated in a case study on 104 customers for a grid with high-penetration of low-carbon technologies. Results show sufficiently accurate forecasts at the customer level and especially at the aggregated level, even at low participation rates.

Content

DDC Class

600: Technology