This paper prescnts the results of the Electric Power Research Institute Short Range Forecasting Project (EPRI-SRF) performed by the Load Forecasts Department, Economics and Forecasts Division of Ontario Hydro, Ontario, Canada.
In this study a variety of short-range forecasting techniques are applied to Ontario Hydro monthly data on total system energy demand. These techniques are available in a software package (FORECAST MASTER) developed for EPRI by two consultants-Scientific Systems, Inc. (SST) and Quantitative Economic Research, Inc. (QUERI).
The methods used for this study were the univariate Box-Jenkins method, the multivariate state-space method, Bayesian vector autoregression and autoregressive econometric regression.
A comparison of the models developed show that the econometric modcls perform the best overall. The state-space models are more suitable for very short-term (one-step ahead) forecasts. Although the Box-Jenkins method has the advantage of simplicity in terms of estimation and data requirement, its performance was not as good as that of the others. Bayesian vector autoregression results indicate that this program needs some modification for monthly data.
KEY WORDS
Electricity demand Box-Jenkins State-space Bayesian vector autoregression
In 1985, Ontario Hydro was approached by the Electric Power Research Institute (EPRI) to participate in their short-range forecasting project (EPRI-SRF). The purpose of the project was, mainly, to help utilities in developing short-range forecasting models through (1) developing software incorporating various state-of-the-art forecasting methodologies, and (2) providing assistance to a few selected utilities to test the software to smooth out problems both in using it and in applying the methodologies to their relevant areas of interest. Ontario Hydro was one of the testing utilities. We used the software (FORECAST MASTER) to model and forecast Ontario Hydro total system energy demand. This paper presents our experience and the results of the models.
Four main forecasting methods were used. Three of these methods, the autoregressive econometric, the state-space and the ridge vector autoregression methods were multivariate. The