Wind Data Interpolation And Long Term Data Prediction Using ANNs Based Method

Wind energy assessment plays an important role in the preliminary development of wind power project.The task of wind energy assessment is to estimate wind resources that represent long-term conditions at a candidate site according to short-term wind data series and data from reference stations,thus determining feasibility of a wind power project.In real practice,the short-term wind data measurements are of significant importance.Whereas,missing records and failure data are likely to happen in condition that the mast being exposed to adverse climate conditions or suffering from instrument failure.Therefore in the first place,wind data measurements are expected to be examined thoroughly with due interpolation to be implemented if needed.Due to the interannual variability of wind speed,thedata measured over one or two years are insufficient to reflect the average wind conditions in the wind farm.Domestically to handle this,representative year correction is the prevailing way.The long-term wind resource that represents the average conditions at a candidate site is then obtained through data correction.However,as is often the case at abroad,a method that is commonly known as the Measure-Correlate-Predict(MCP)is analternative approach,which produces long-term series of wind data through hindcasting.As is seen above,Interpolation and long-term correction of wind data measurements are the foundation of wind resource assessment,and the accuracy of them can somehow influence economic efficiency of a wind farm.In view of the deficiency of the current interpolation and correction method at home,researches on new way of interpolation and long-term wind data MCP hindcasting based on ArtificialNeural Networks(ANNs)have been conducted in this thesis.As a result,the interpolation accuracy has been improved and the correction error eliminated,and the main contentsof this thesis are listed as follows.Out of the advantage of ANNs in fitting multi-input and non-linear relations,wind data measurements interpolation using BP networks and ELM method are modelled.Meanwhile,a simple nonlinear function is used for model validation proving the advantage of the BP and ELM algorithm.The BP and ELM based interpolation model has an advantage over the linear interplolation when this approach is apllied to a complex terrain mast and a simple terrain mast.For example,as is the case with interpolation within a same mast of simple terrain,RMSE is brought down by nearly 30%,R2increases from 0.9973 to 0.9990;When it comes to the case with interpolation according to reference stations,RMSE is reduced by 4%,and R2 rises by 1.9%.In condition of interpolation within a same mast of complex terrain,RMSE is diminished by 40%,and R2 is improved from 0.9881 to 0.9960.In case of interpolation according to reference stations,the BP and ELM based approach has a RMSE decrese of 27.4%and 13.7%respectively in the 30m and 70m layer.The BP algotithm is optimized through Genetic Algorithm for a better initial weight and threshold value.And a correlation model based on GA-BP and the ELM is formed for the MCP procedure.This model of higher accuracy not only allows a multi-input involving wind speed,wind direction and other meteorological parameters,etc,but also eliminates the correction error caused by the prevailing way of representative year.The accuracy of both GA-BP and ELM based MCP is higher with comparison to a linear approach.As is shown by the calculation,for a simple terrain mast,BP has a smaller RMSE of 3.6%,while ELM 4%,and GA-BP 5.4%.As for a complex terrain mast,BP has a smaller RMSE of 7.9%,while ELM 11.3%and GA-BP 12.8%.In this thesis,anANNs based model of higher accuracy for interpolation is formed.When data interpolation is influenced by the difference lies in measuring height inpterpolation method and landform,errors of ANNs based method are smaller than that of a linear method.Meanwhile,the ANNs based method forlong term series hindcasting acquires the long term wind data by hour or by month,which throws light on a new way of preparing the wind data.