Risk adjusted forecasting of electric power load

Load forecasting of energy demand is usually focused on mean values in related statistical models and ignores rare peak events. This paper provides Extreme Value Theory analysis of the peak events in electrical power load demand. It estimates risk of the peak events by combining forecast of the mean with extreme value modeling of distribution tail. The approach is demonstrated for electric load demand data for a US utility. The problem is to find the forecast margins that keep the risk of demand exceeding forecast plus the margin to one event per year. The long tail model of the peak events is more accurate and yields 50% larger margin compared to the normal distribution model. These results show that the long tail behavior of the forecast errors must be taken into account when trying to keep outage risk low.