Short-term Load Forecasting By Multi-feature Iterative Learning Based on ISFS And XGBoost

Abstract

With the continuous development of smart grid and demand response technology, the electrical load gradually takes on an elastic, flexible, uncertain and controllable quality. In order to fully excavate the potential information behind the power load in smart grid and make use of it, the study of load feature analysis and load forecasting appear to be particularly important. Considering massive related data, machine learning algorithm based on big data analysis is the current mainstream method of establishing a forecasting model. It deeply mines the mapping relation between features and load. Feature selections on model training will directly affect the accuracy of short-term load forecasting. For making the most of massive data to improve the effect of feature selection, this paper proposes a short-term load forecasting method by multi-feature iterative learning based on ISFS (Improved Spanning-tree Forward Selection) and XGBoost (eXtreme Gradient Boosting). Under the framework of iterative learning, the proposed method uses ISFS algorithm to make better feature selection successively by iterations. And XGBoost algorithm evaluates each feature selection by cross validation results of training data set, thus precisely finding out the optimal multi-synergistic relationships among impact features and building differentiated models with distinct feature subsets. The method accumulates information gain by re-studies from the iterative load forecasting results, manages to improve the training effect and reduce the load forecasting errors step by step. The experimental results show that the proposed method has higher load forecasting accuracy compared with other typical methods.