Probability distributions of short-term electricity peak load forecasts

Electric companies schedule generator maintenance so as to equalize the risks of capacity shortfall. Distributions of peak loads for given future weeks, months and seasons provide information on the probability that a critical load level will be exceeded in the given time period. Day-to-day fluctuations in peak loads were assumed to be directly caused by weather variables whose distributions vary by week throughout the year. Forecast distributions of weekly peak loads were computed by non-parametric simulation and re-estimation for three regression models: (1) weekly peak loads dependent on time trend and dummy variables, (2) weekly peak loads dependent on socioeconomic and weather variables, and (3) a daily version of (2). Comparisons of actual peak loads against within-sample and post-sample forecast distributions of peak loads showed that (1) and (2) were statistically equivalent but (3) was significantly worse.