Discussion on ‘Electrical load forecasting by exponential smoothing with covariates’

Abstract

Rainer Gob, Kristina Lurz and Antonio Pievatolo (hereinafter GLP) address a very important issue in power systems management—load forecasting. Generally, load forecasting is concerned with the accurate prediction of the electric load (or demand) for specific geographical locations and over the different periods of the planning horizon. However, as in the discussed paper, the quantity of interest is usually the hourly total load, and the planning horizon is short-term—it ranges from a few minutes to a few weeks. Short-term load forecasting has become increasingly important since the rise of competitive energy markets. Electric utilities (called ‘energy vendors’ by GLP) are the most vulnerable as they typically cannot pass costs directly to the retail consumers. When electricity sectors were regulated, utility monopolies used short-term load forecasting to ensure the reliability of supply (prevent overloading, reduce occurrences of equipment failures, etc.). Nowadays, the costs of over-contracting/under-contracting and then selling/buying power in the balancing (or real-time) market are typically so high that they can lead to huge financial losses of the utility and bankruptcy in the extreme case. Load forecasting has become the central and integral process in the planning and operation not only of electric utilities (as GLP clearly point out) but also of energy suppliers, system operators and other market participants. A variety of methods and ideas have been tried for load forecasting, with varying degrees of success. Following Weron [1], GLP classify them into two broad categories: (i) statistical approaches, including similar-day (or naive), exponential smoothing, regression and time series methods; and (ii) artificial intelligence-based techniques. Among them, exponential smoothing stands out as a simple yet powerful approach, whereby the prediction is constructed from a weighted average of past observations with exponentially smaller weights being assigned to older observations. More complex variants—such as the Holt-Winters method—have been developed to model time series with seasonal and trend components [1, 2]. Application of exponential smoothing to hourly electricity load data requires further generalization to accommodate the prevailing seasonalities (daily, weekly and possibly annual) and weather-related exogenous variables (called ‘covariates’ by GLP). This discussion paper offers comments in three sections. The first section discusses the model and its position within the exponential smoothing literature on load forecasting. The second section comments on the empirical part of the paper. The final section offers recommendations for further research.