A two-stage random forest method for short-term load forecasting

Abstract

Machine learning methods are the main stream algorithms applied in short term load forecasting. However, typical machine learning methods consisting of Artificial Neural Network (ANN) and Support Vector Regression (SVR) have deficiencies hard to overcome, such as easy to be trapped in local optimization (for ANN) or hard to decide kernel parameter and penalty parameter (for SVR). On the other hand, grey relational analysis is an effective method to select proper historical data as training set for training machine learning models. But it is not comprehensive and accurate enough. In this paper, a new two-stage hybrid algorithm aimed to solve these two problems is proposed. Random Forest (RF) method is introduced as the machine learning method, which will not cause overfitting problem and parameters are easy to be tuned. Furthermore, Grey Relational Projection (GRP) is introduced to select similar historical data to train random forest models. The final forecasting results based on real load data prove this new two-stage method performs better than the other two common methods.