Wind Power Forecasting Based On Machine Learning

In recent years,wind power has developed rapidly around the world to cope with the contradiction between economic development,energy demand and environmental protection.Based on achieving the goal of green and sustainable development,China is accelerating the development and utilization of wind power,and has taken the lead in the world.With the increase of wind power project,China urgently needs to improve the utilization efficiency.However,affected by natural factors,wind power has randomness,volatility and complexity,which affect the grid connected consumption and utilization efficiency of wind power.Under the existing equipment and technical conditions,improving the level of wind power forecasting is one of the effective methods to overcome these adverse factors,which is conducive to improving the accuracy of wind energy resource assessment,optimizing production and maintenance plans,and improving the peak load regulation of power system and the real-time control ability of wind turbines.The traditional prediction methods based on physics and statistics are difficult to deal with complex nonlinear data,and the emerging machine learning methods try to make up for this deficiency,which has been widely valued by the academic and business circles.On the basis of literature research,this paper proposes improved data processing methods and forecasting models to solve the shortcomings and problems of existing wind power prediction methods.The main research contents and innovations are as follows:(1)Predictability analysis and distribution estimation.The predictability of wind speed is the basis of wind power forecasting,but there are few related studies at present.A point-topoint similarity comparison method is designed to prove that the historical data contains data segments similar to the future data.Through logic analysis and statistical analysis,this study proves that the wind speed data contains predictable trend and periodic characteristics.To solve the problem that the kernel function is difficult to select in the existing nonparametric distribution estimation methods,this paper constructs a kernel function which peak value and integral value are equal to 1,and analyzes the influence of the peak value and shape of the kernel function on the distribution estimation results.The empirical results show that the wind speed distribution is changing under different time windows,and the improved kernel function can improve the estimation accuracy of wind speed distribution.(2)Based on the analysis of wind speed distribution,this paper attempts to improve the accuracy of wind speed distribution to serve the wind energy resource assessment.Most of the existing studies use parameter estimation methods to predict the parameter changes of wind speed distribution function,and the accuracy is low.The data quantization method of nonparametric estimation result is designed to solve the predictability problem of it.This paper presents a multi-channel integrated prediction method based on improved differential evolution optimization algorithm and hybrid machine learning model.The results show that the proposed optimization algorithm and prediction method are suitable for the prediction of wind speed distribution,and the prediction accuracy is better than the comparison model.(3)To solve the problem that the cumulative error of wind speed medium-term prediction increases rapidly with the increase of prediction step,this paper designs an improved "data feature extraction-horizontal multi-step prediction-vertical integration" prediction model,which places the multi-step prediction process inside the model to reduce the cumulative error.To avoid the influence of artificial selection,a new method is designed to determine the decomposition number of Singular Spectrum Analysis.A data set expansion method based on random number is designed,which can expand the training data set and improve the generalization ability of neural network.The results show that the proposed model can avoid the rapid increase of cumulative error in the prediction process.(4)To solve the problem that the new wind farm cannot use enough historical data to train the machine learning prediction model,this paper proposes a dual-meta pool prediction method,which realizes the prediction of small sample data by learning a large amount of data knowledge contained in adjacent wind farms.This paper defines meta-data that can reflect data knowledge,designs an improved unsupervised classification method for data element classification,and uses neural network to construct the mapping relationship between meta-data and their labels after classification.In order to realize the prediction of meta-data,this paper proposes metamethod which includes the method of Hilbert data spatial features construction and neural network,and designs a dual-meta pool prediction method based on the mapping relationship between the meta-data and meta-method.The results show that the proposed method can be used for short-term wind power prediction,and has higher prediction accuracy than other comparison models.(5)In order to improve the accuracy and speed of ultra-short-term power prediction of wind turbine,this paper proposes an online dual-mode transfer learning prediction method based on multi-source data fusion.By canceling the weight calculation unit and adding two data buffers,this method realizes the system-wide update and high-speed data prediction.To enrich and fuse multi-source data,this paper proposes new time trend quantification method to adds time trend data,and uses neural network to construct the mapping relationship between multi-source data and wind turbine power.The results show that the new method including data transfer learning and method transfer learning is suitable for real-time online prediction.This study is based on actual measurement data,trying to solve the practical problems of wind power forecasting by improving data processing and forecasting methods,to provide valuable reference for the development and utilization of wind power energy in China.