Research On Wind Power Forecasting Approach Based On Ensemble Learning

Since the large-scale grid connection of wind power will bring a very big impact to the power grid and affect the stable operation of the power grid,the research on wind power forecasting is of great significance for the safe and stable operation of the power grid and the development of new energy.In order to improve the accuracy and stability of Wind Power Forecasting(WPF),an ensemble learning WPF model based on SOM clustering and K-fold cross-validation optimization is proposed.BMA and ANFIS are used as the elementary learning tools for ensemble learning,and the simulation experiments prove that the proposed approach has better prediction effect.This paper studies the following:(1)Introducing the background and significance,the research status of WPF field at home and abroad,and summarize the problems of different prediction methods.The ensemble learning algorithm is introduced,and the main contributions and structural arrangements of this paper are given.(2)From the perspective of the principle of wind power generation,the main factors affecting wind power are meteorological factors,and the effects of wind speed,wind direction and temperature on wind power generation are analyzed.The SCADA data of wind farms are introduced,and various evaluation indicators are proposed to judge the accuracy and generalization performance of WPF models.(3)Preprocessing the data,predicting the single model of wind power,and verifying the optimization effect of Pearson similarity processing on single model prediction.Firstly,the SCADA raw data is preprocessed,and the Gaussian fitting power curve is used to eliminate the abandonment data.Then,according to the Pearson correlation analysis,the optimal sampling interval is selected,and the weights of wind speed,wind direction and ambient temperature are calculated.Finally,BPNN,SVM and RBFNN are used to predict the single model.The simulation results show that the preprocessing has obvious optimization effect on single model prediction.(4)For the problem of single model prediction instability and generalization performance,a processing method for training subset diversity and a new Bayesian Model Averaging(BMA)ensemble learning WPF method are proposed.First,the training set is split into multiple training subsets with the same data distribution using a combination of SOM clustering and K-fold cross-validation.Then,using BMA as a meta learner,integrate the output of three single model predictions for WPF.Finally,the simulation experiment proves that the WPF method proposed in this chapter has excellent precision and stability.(5)Selecting a new meta-learner-Adaptive Network-based Fuzzy Inference System(ANFIS)for WPF and compare it with simulation.The simulation results show that the prediction accuracy and stability of ANFIS-EL are higher than that of the primary learner.In addition,the difference between the prediction results of ANFIS-EL and BMA-EL shows that the meta-learning device of Stacking-type integrated learning is different,the optimization effect is different,and the prediction time is different,so different models can be selected according to actual conditions.