LAUSR

Abstract Electric power utilities need accurate load forecast model for reliable and efficient planning and operation of the grid. For short-term load forecast such as day-ahead case, data-driven mathematical models have been traditionally studied with various methods such as linear regression, time-series analysis, support vector machines and artificial neural networks. With large amount of historic data used for training the models, experimental approaches have usually attempted to show incremental improvement in accuracy. In experience, however, it has been observed that improvement in accuracy is limited to around a few percent of errors, which we consider should be caused by certain reasons. This motivated us to develop a theory to quantify the amount of historic data necessary to achieve a certain level of accuracy. This paper addresses the issue of quantifying the trade-off between model data requirement and accuracy. Using results of an examination performed on data from an electric utility, we demonstrate a novel method and mathematical criteria to judge the trade-off.