Locally Weighted Ensemble Learning For Wind Speed Prediction

With the gradual depletion of traditional energy,the development and utilization of renewable energy have become a hot issue.Wind energy,one of the fastest growing renewable energy,is becoming more and more important.Wind power generation as the most important utilization of wind energy,is a hot issue in the current research of new energy power industry.However,because of the characteristics of random fluctuation and intermittent,the uncertainties of power system operation will increase when wind farms are connected to the grid.If more accurate wind speed power is obtained,it will be conducive to develop a scheduling plan by dispatching department timely,alleviate the adverse impact caused by intermittency characteristic of wind power,and ensure the security and stability of the power grid.Wind speed prediction plays an important role in wind power penetration.In this paper,wind speed prediction in wind farm is the main research content,which is shown as follow:Firstly,Local ensemble learning is proposed for accurate short term wind speed prediction based on neighboring samples.At different time,the variation of wind speed is different.Local learning is a class of relatively new technique to train highly nonlinear functions.In this paper,we introduce ensemble learning into local learning,and design a local ensemble learning algorithm for combating the variation of wind speed.The proposed algorithm search the nearest neighbors of each test sample through K nearest neighbor algorithm and train multiple base models with these neighboring samples.Thus each test sample may be predicted with different models.Then the predicted values from these models are combined with a certain strategy.The proposed algorithm is tested in short-term wind speed prediction.Two datasets are collected from Ningxia and Jilin,China,respectively.Significant reduction of prediction errors is observed from the experimental results.Secondly,Locally Weighted Ensemble Learning is proposed for regression.The weight varies with sample,which is realized by introducing soft-max function into the objective function.The introduction of soft-max function makes the objective function ill-posed.Regularization is a way to overcome this problem.L21-regularization,LF-regularization and Laplacian-regularization are applied to the objective function.The proposed method is evaluated on eight UCI datasets and Jilin wind farm.Compared with single models,constant weight ensemble models,dynamic weighted ensemble models and Adaboost,significant result is achieved.From the experiments,the convergence of LWE is fast.Thirdly,stacked-locally weighted ensemble method is proposed for wind speed prediction.Wind speeds are typical of the time series;the neighboring wind speeds are more relevant.The previous time prediction is treated as one attribute in the prediction of next time wind speed.The architecture of SLWE not only considers the diversity but also the accuracy.The proposed method is tested on wind speed datasets from Jilin,China.Experimental results show the effectiveness of stacked-locally weighted ensemble method.