End-to-end Learning for Hierarchical Forecasting of Renewable Energy Production with Missing Values
Abstract
Power systems feature an inherent hierarchical structure. Ensuring that forecasts across a hierarchy are coherent presents an important challenge in energy forecasting. In this context, proposed reconciliation or end-to-end learning approaches assume coherent historical observations by construction; this assumption, however, is often violated in practice due to equipment failures. This work proposes an end-to-end learning approach for hierarchical forecasting that directly handles missing values. First, we show that a class of off-the-shelf machine learning algorithms already leads to coherent hierarchical forecasts. Next, we describe a conditional stochastic optimization approach based on prescriptive trees for end-to-end learning with missing values, without imputation or disregarding of quality observations. We validate the proposed approach in two case studies comprising 60 wind turbines and 20 photovoltaic parks, respectively. The empirical results show that end-to-end learning outperforms twostep reconciliation approaches and that the proposed solution mitigates the adverse effect of missing values.
Origin : Files produced by the author(s)