Improvement of the Short Term Load Forecasting Through the Similarity Among Consumption Profiles

In order to achieve high quality standards in electrical power systems, utility companies rely upon load forecasting to accomplish critical activities such as optimal dynamic dispatch and smart performance in the power wholesale market. Several works propose hybrid intelligent forecasting models to deal with the dynamic and non-linear characteristics of the load at a relatively high computational cost. While such approaches give emphasis to the forecasting itself, this paper presents a procedure to detect similarities among distinct consumption profiles. Empirical results show that similar profiles share similar sets of relevant predictors. As finding similarities among profiles is less costly than finding the set of relevant predictors from scratch, a new parameter selection method is proposed. Such method is employed to build some neural forecasters with a marked improvement in the learning time.