Comparing Univariate and Multivariate Methods for Probabilistic Industrial Load Forecasting

Abstract

Due to their large usage of electricity, industrial factories stand out from domestic and commercial utilities for the enormous potential in providing services to power systems. Accurate energy consumption forecasts are required in order to exploit this outstanding capability, without incurring into operational mistakes. In this context, system operators prefer probabilistic load forecasts when dealing with decision-making processes, in order to fully account economical and technical risks. This paper tackles probabilistic industrial load forecasting from a dual point of view: active and reactive power forecasting. The strong correlation between active and reactive powers suggests to develop the proposed forecasting method by modeling the interaction effects between the variables; we investigate the convenience of such approach using both univariate and multivariate methods, i.e., quantile regression forests and vector autoregressive exogenous models, trained with actual data registered in an Italian factory. The results are compared to a naïve benchmark and a regression bootstrap benchmark.