A Short-Term Forecasting Method for Photovoltaic Power Based on Ensemble Adaptive Boosting Random Forests

Accurate and reliable forecast of Photovoltaic (PV) power is crucial for power system dispatch and control. In this paper, a novel short-term PV generation forecast approach based on ensemble adaptive boosting random forests (EABRF) is proposed. The approach sets the weight distribution for training samples and bootstrap sampling is applied. By this means, different sub-training data sets with weights are generated to train the decision trees. The output of random forests is that the average value of each decision tree. Further, by calculating the error rate and assigning corresponding weight to each random forests, the weighted result of random forests can be obtained. The distinguishing features of the proposed approach are as follows: (a) The approach reduces the risk of overfitting by using the idea of ensemble learning; (b) The proposed approach can plenty mine the information in multidimensional features; (c) It has good generalization and robustness by combining adaptive boosting with random forests. PV power forecasting has been conducted in four PV plants using the proposed approach with the historical PV power and numerical weather prediction (NWP) as the inputs. The results illustrate that the proposed method is effective for forecasting of PV power generation with satisfactory forecast accuracy.