Exploration and Comparison of New Methods for Electric Demand Forecasting

Abstract

A variety of time series forecasting methods were tested for the prediction of state-wide monthly residential electrical sales per customer, in a formal experimental design across ten states in the U.S. Models were fitted to historical data from 1962 through 1977, and evaluated for their performance over the following three-year period from January, 1978 through December, 1980. Rolling forecasts were generated on a monthly basis through the evaluation period to obtain 36 one-month forecasts, 35 two-month forecasts, etc. and were reduced to a set of summary statistics for different horizons and averaging methods. The forecasting models included several econometric techniques and several pure time series techniques. The conclusions were that there were gains in accuracy for structural models (use of exogenous variables), for dynamic models (involving lagged dependent variables), and for adaptive models. It was further concluded that the systematic use of model selection criteria such as the Akaike Information Criterion, and of diagnostic tests, led to improved accuracy in the accepted models.