Hierarchical load forecasting
Ines Dormoy and Louis Gautier
Forecasting energy load at several scales is a crucial task for grid managers and utilities, particularly with the rise of intermittent renewable energy sources demanding a heightened level of planning. In this project, we explore ways to predict energy demand at several levels of hierarchy in a grid, from individual meters to large substations serving entire neighborhoods. A desirable property in this context is that the sum of the load time series at the bottom of the hierarchy (e.g., individual meters) matches the load at higher levels. We propose two methods to achieve this hierarchical load forecasting objective: a two-step approach consisting of reconciling base learners and a deep learning-based end-to-end pipeline. We benchmark several design choices for these two methods and analyze whether they improve accuracy at various levels of the hierarchy. Finally, we compare them in terms of robustness and tractability on real massive datasets.