Estimation of individual prediction reliability using error analysis applied to short-term load forecasting problem

This work describes the methodology to create a reliability estimate for individual predictions in regressions. This estimate is defined as a binary variable which indicates if the regression prediction error of an individual unseen observation is likely to be critical or not, according to a meaningful criterion previously defined by the regression model user. The approach is based on the construction of a model to separate these two classes of error. The method was evaluated on sixteen experiments applied to short-time load forecasting regression problem using eight databases from ISO New England. In these experiments, the models for pattern recognition were built as ensembles composed of three classification models: K-Nearest Neighbors, Artificial Neural Network Committee Machine, and Support Vector Machine. The obtained results showed that the Ensemble Classifiers were able to detect critical error cases.