Building’s hourly electrical load prediction based on data clustering and ensemble learning strategy

Abstract

The short-term power load forecasting of buildings plays a key role in peak demand management and power generation scheduling. Many studies have shown that ensemble learning strategies are more accurate and feasible in practice than individual predictive models. However, the common way processing massive data is still traditional, that is, training and testing them as a whole, ignoring data’s inter law and pattern difference, which probably hinders the further improvement of forecasting effect. In this study, the characteristics of building data are identified and classified before model construction. Fully considering the diversities of model structure and parameters, an adaptive ensemble learning strategy is proposed for short-term building electrical loads forecasting. The unsupervised K-means clustering and supervised KNN (K-nearest neighbors) classification methods are combined for data clustering. Total eleven different sub-learners are applied for ensemble learning and a variety of intelligent optimization algorithms are used for parameter adjusting. The overall prediction accuracy is derived through a fusion module. Two sets of short-term electrical load data of actual buildings are used for model verification. The results show that the data clustering based ensemble learning strategy has better forecasting accuracy compared to eleven individual models and previous reported predictions. The superior prediction accuracies in different two cases also verify the adaptability and generalization abilities of the proposed ensemble method.