Robust short-term load forecasting

Analyzing the stochastic characteristics of electric consumption series in many countries shows the presence of atypical observations or outliers. Outliers are deviant data points that do not follow the model of the majority of observations. They significantly degrade the accuracy of conventional day-ahead estimation and forecasting methods even if they are present in a very low fraction. Thus, both their identification and separate treatment are important in practice. In this paper, we evaluate by a comparison analysis the performance of some robust estimation methods. Namely, the robust filtered S- and filtered τ-estimators, the minimum covariance determinant (MCD), and the 3-σ rejection rule. The performance of these methods has been evaluated on the French electric demand in terms of forecasting accuracy. The sophisticated robust methods exhibit the best reliability if compared to non-robust or basic robust methods. Practical load forecasting confirms the necessary good theoretical tradeoff between robustness and efficiency under the nominal model.