Empirical Comparison of Prediction Methods for Electricity Consumption Forecasting

Recent years have seen an increasing interest in providing accurate prediction models for electrical energy consumption. In Smart Grids, energy consumption optimization is critical to enhance power grid reliability, and avoid supply-demand mismatches. Utilities rely on real-time power consumption data from individual customers in their service area to forecast the future demand and initiate energy curtailment programs. Currently however, little is known about the differences in consumption characteristics of various customer types, and their impact on the prediction method’s accuracy. While many studies have concentrated on aggregate loads, showing that accurate consumption prediction at the building level can be achieved, there is a lack of results regarding individual customers consumption prediction. In this study, we perform an empirical quantitative evaluation of various prediction methods of kWh energy consumption of two distinct customer types: 1) small, highly variable individual customers, and 2) aggregated, more stable consumption at the building level. We show that prediction accuracy heavily depends on customer type. Contrary to previous studies, we consider the consumption data granularity to be very small (i.e., 15-min interval), and focus on very short term predictions (next few hours). As Smart Grids move closer to dynamic curtailment programs, which enables demand response (DR) events not only on weekdays, but also during weekends, existing DR strategies prove to be inadequate. Here, we relax the constraint of workdays, and include weekends, where ISO models consistently under perform. Nonetheless, we show that simple ISO baselines, and short-term Time Series, which only depend on recent historical data, achieve superior prediction accuracy. This result suggests that large amounts of historical training data are not required, rather they should be avoided.