Probabilistic short-term load forecasting with conditional mean-variance and quantile regression models

Abstract

For the day-ahead density forecasting of electricity load, this paper proposes the combination of the autoregressive moving average (ARMA) model and the generalized autoregressive conditional heteroskedasticity (GARCH) model, with both of them admitting exogenous inputs. This composite structure on the conditional mean and variance is referred to as the ARMAX-GARCHX model. As an alternative to its estimation by means of log-likelihood maximization, approaches based on iterative least-squares (ILS) and nonlinear least-squares (NLS) are considered. Apart from the ARMAX-GARCHX model, quantile regression models (QRMs) are also tested in forecasting where a wide range of quantiles are separately modeled to approximate a density. Phase currents of several low voltage transformer cables from the Netherlands are forecasted to compare the performances, and as the probabilistic evaluation criterion, the continuous ranked probability score is used. As an outline of the results, the ARMAX-GARCHX model outperformed QRMs and among its estimation techniques, the likelihood-based approach had the best performance, though the differences in the errors are often minor. Thus, owing to its computational simplicity, the ILS solution can be a valuable option when processing large batches of data in practice.