Short-term load forecasting with a hybrid clustering algorithm

Abstract

Load forecasting is an important part of the operational procedure for a power system and a considerable amount of research effort has been expended on the development of accurate prediction methodologies. The electrical load series leads the field for the construction and application of state-of-the-art forecasting models, especially those based on artificial intelligence. The hybrid models, which are developed using a clustering algorithm to group data with similar characteristics and a function approximation to capture the underlying characteristics of each cluster of data, form a special class. For the majority of clustering algorithms, clusters are formed using some distance measure, thus identifying each cluster as a group of data allocated closely together. The clustering scheme that is developed generates clusters that are described by the same linear model. A demonstration of the proposed methodology is performed for the one-step ahead forecasting of load data from the Californian and the New York state power systems. The analysis of the forecasting results showed that the proposed algorithm was able to reduce the forecasting error by 7.5% and 9%, respectively, for the two data sets, compared to a neural network developed using the traditional load forecasting methodology.