Probabilistic load forecasting with generative models

Abstract

Electric load forecasting is a central step for economic actors to plan supply purchases and for energy system operators to control networks. The marginal gain in forecast accuracy potentially leads to important resources saving for power companies involved. With this in mind, researchers in the field of power system shifted from point forecast to probabilistic forecast in order to produce more informative and reliable predictions, then to characterize uncertainty tied to the forecast. In turn, this work explores the introduction of information from an electrical grid topology into the probabilistic forecast framework, hoping to further improve load forecasters. Building upon normalizing flows, we apply graphical normalizing flows to the field of power system. Normalizing flows are a class of generative models that create a mapping between the distribution of interest and a known distribution. Graphical normalizing flows allow to add inductive bias by taking into account prescribed dependencies between random variables from the targeted multivariate distribution. In this particular study, we are interested in the forecast of day-ahead electric load from different connected zones. Experiments are conducted on a actual case study, hourly aggregated electric load collected over 14 years in all 8 states of New England, USA. After analysing the data, we investigate the performance of graphical normalizing flows when proposing various topology for the electrical network. We finally compare the results to other generative models, namely autoregressive normalizing flow and variational autoencoder. We observe tuned and trained graphical normalizing flow forecaster achieves slightly better result on various metrics than all other models globally. We show that GNF with well-identified independence introduces an inductive bias sufficient to improve model distributions and scenarios generated.