Research On Wind Power Forecasting Based On Deep Learning

The intermittent of wind power posing serious influences to the safe,stable and economical operation of power system,it is the main challenge of large-scale wind power integration.Wind power prediction is one of the necessary tools to solve this problem.The main problem of wind power prediction is that the accuracy is not high enough.There are two main reasons for the problem: on the one hand,wind power forecasting model regularly use one-to-one mapping and shallow structure;On the other hand,the training samples tend to be few dimensions,small scale,and without a strict cleaning.Thus,learning ability of model is insufficient,it is difficult to adapt to different wind scenarios,which greatly limits the improvement of the prediction accuracy.In this paper,a method of wind power prediction based on Stacking Denoising Auto-Encoder(SDAE)is proposed,which has multi-NWP(numerical weather prediction)as inputs and multi-wind turbines generation as output.Main contributions include:1)Wind power forecasting data cleaning method analysisQuality of model data for wind power forecasting are firstly analyzed,then two kinds of imputation algorithm are proposed to solve missing and abnormal situation of measured wind data.One kind of filter method based on pitch angle upper bound line is proposed for selection normal measured wind power data.The precision of KNN imputation algorithm is higher,more suitable for subsequent modeling work;the proposed filtering method can quickly and accurately select measured power data in the normal running state.2)Multi-NWP correction method based on SDAEThe error pattern of NWP data is analyzed,and a method for multi-NWP correction based on SDAE is proposed.Established model is a three hidden layer SDAE network which input is multi-NWP and output is wind speed of each wind turbines.Three comparison model are set up.After comparison and analysis,it is found that the multi-NWP correction model based on SDAE has a strong learning ability and does not need to build model for each month and wind turbine unit.The correction performance is stable and the correction precision is high.3)Wind power prediction method based on corrected multi-NWP wind speed andSDAEThe advantages and difficulties of multi-dimensional input and multi-dimensional output mapping structure are analyzed.A wind power forecasting method based on corrected NWP speed and SDAE are present.The forecasting model is a three hidden layer deep network with corrected multi-NWP as input and multi-wind turbine power as output.Eight conventional structure wind power forecasting models are setted as comparison model.Comparative analysis shows that NWP correction has a significant effect on the improvement of power prediction accuracy.Multi-NWP input will help enhance accuracy of wind power prediction.Three hidden layer SDAE network has a better performance than shallow network model.