The majority of the load forecasting literature has been on point forecasting, which provides the expected value for each step throughout the forecast horizon. In the smart grid era, the electricity demand is more active and less predictable than ever before. As a result, probabilistic load forecasting, which provides additional information on the variability and uncertainty of future load values, is becoming of great importance to power systems planning and operations. This paper proposes a practical methodology to generate probabilistic load forecasts by performing quantile regression averaging on a set of sister point forecasts. There are two major benefits of the proposed approach. It can leverage the development in the point load forecasting literature over the past several decades and it does not rely so much on high-quality expert forecasts, which are rarely achievable in load forecasting practice. To demonstrate the effectiveness of the proposed approach and make the results reproducible to the load forecasting community, we construct a case study using the publicly available data from the Global Energy Forecasting Competition 2014. Compared with several benchmark methods, the proposed approach leads to dominantly better performance as measured by the pinball loss function and the Winkler score.Index Terms-Electric load forecasting, forecast combination, pinball loss function, prediction interval (PI), probabilistic forecasting, quantile regression, sister forecast, Winkler score.

Temperature plays a key role in driving electricity demand. We adopt "recency effect", a term originated from psychology, to denote the fact that electricity demand is affected by the temperatures of preceding hours. In the load forecasting literature, the temperature variables are often constructed in the form of lagged hourly temperatures and moving average temperatures. Over the past decades, computing power has been limiting the amount of temperature variables that can be used in a load forecasting model. In this paper, we present a comprehensive study on modeling recency effect through a big data approach. We take advantage of the modern computing power to answer a fundamental question: how many lagged hourly temperatures and/or moving average temperatures are needed in a regression model to fully capture recency effect without compromising the forecasting accuracy? Using the case study based on data from the load forecasting track of the Global Energy Forecasting Competition 2012, we first demonstrate that a model with recency effect outperforms its counterpart (a.k.a., Tao's Vanilla Benchmark Model) in forecasting the load series at the top (aggregated) level by 18% to 21%. We then apply recency effect modeling to customize load forecasting models at low level of a geographic hierarchy, again showing the superiority over the benchmark model by 12% to 15% on average. Finally, we discuss four different implementations of the recency effect modeling by hour of a day.

Although combining forecasts is well-known to be an effective approach to improving forecast accuracy, the literature and case studies on combining load forecasts are relatively limited. In this paper, we investigate the performance of combining so-called sister load forecasts, i.e. predictions generated from a family of models which share similar model structure but are built based on different variable selection processes. We consider eight combination schemes: three variants of arithmetic averaging, four regression based and one performance based method. Through comprehensive analysis of two case studies developed from public data (Global Energy Forecasting Competition 2014 and ISO New England), we demonstrate that combing sister forecasts outperforms the benchmark methods significantly in terms of forecasting accuracy measured by Mean Absolute Percentage Error. With the power to improve accuracy of individual forecasts and the advantage of easy generation, combining sister load forecasts has a high academic and practical value for researchers and practitioners alike.

We present a lasso (least absolute shrinkage and selection operator) estimation based methodology for probabilistic load forecasting. The considered model can be regarded as a bivariate time-varying threshold autoregressive(AR) process for the hourly electric load and temperature. The joint modeling approach directly incorporates the temperature effects and reflects daily, weekly, and annual seasonal patterns and public holiday effects. We provide two empirical studies, one based on the probabilistic load forecasting track of the Global Energy Forecasting Competition 2014 (GEFCom2014-L), and the other based on another recent probabilistic load forecasting competition that follows the similar setup as GEFCom2014-L. In both empirical case studies, the proposed methodology outperforms two multiple linear regression based benchmarks from a top 8 entry of GEFCom2014-L.

No abstract