Short-term electricity load time series prediction by machine learning model via feature selection and parameter optimization using hybrid cooperation search algorithm

Reliable load time series forecasting plays an important role in guaranteeing the safe and stable operation of modern power system. Due to the volatility and randomness of electricity demand, the conventional forecasting method may fail to effectively capture the dynamic change of load curves. To satisfy this practical necessity, the goal of this paper is set to develop a practical machine learning model based on feature selection and parameter optimization for short-term load prediction. In the proposed model, the ensemble empirical mode decomposition is used to divide the original loads into a sequence of relatively simple subcomponents; for each subcomponent, the support vector machine is chosen as the basic predictor where the real-valued cooperation search algorithm (CSA) is used to seek the best model hyperparameters, while the binary-valued CSA is set as the feature selection tool to determine the candidate input variables; finally, the aggregation of all the submodules' outputs forms the final forecasting result. The presented method is assessed by short-term load data from four provincial-grid dispatching centers in China. The experiments demonstrate that the proposed model can provide better results than several conventional models in short-term load prediction, while the emerging CSA method is an effective tool to determine the parameter combinations of machine learning method.

Automated summary

This paper proposes a practical machine learning model for short-term load prediction based on feature selection and parameter optimization. Experimental results show that the proposed model outperforms several conventional models in short-term load prediction, and the CSA method is an effective tool for determining parameter combinations.