In view of the problems with typical machine learning algorithms (for example artificial neural network (ANN) and support vector machine (SVM))

such as the difficulty in determining the number of hidden units and parameter optimization

a random forest regression method is first introduced to power system load forecast. A new combinatorial algorithm involving two steps is proposed. Firstly

a grey relational judgment matrix is built to characterize the relationship between historical samples and forecasting sample. Secondly

the entropy method is used to determine the weights of all load influencing factors and the weighting matrix is got. Thirdly

the historical samples with bigger grey relation projection values are used to form the training set. After getting the training set

this data set is used to train random forest models. Then

the eigenvectors of the forecasting day are input to the trained model to finish the forecasting process. The real load data of one city in Zhejiang Province is used to test the proposed algorithm

and the results are compared with SVM and random forest method with no improvement made on grey relation projection. The results show that the new combinatorial method has higher precision and robustness than the other two methods. This work is supported by National Natural Science Foundation of China (No. 51277009).