Ensemble learning with time-series clustering for aggregated short-term load forecasting

Load forecasting, as one of the important research areas of the smart grids, spans a wide range of methods, from traditional time-series econometric analyses to different machine learning and, recently, even deep learning approaches. This paper proposes a novel machine learning approach for short-term load time-series forecasting, which utilizes aggregate load clustering with ensemble learning based on the windowing method. Ensemble of base learners, comprised of gradient boosting, support vector machine (SVM) and random forest, is created by stacking models with an "elastic net" linear regression. Models hyper-parameters are fine-tuned using a grid search with cross-validation approach, except for the SVM, where Bayesian optimization is introduced. Features engineering and selection based on the importance analysis is employed, using weather and load time-series data. The mean absolute percentage error is used for verification. Obtained results show that the proposed approach exhibits accurate and robust predictions.