Interpolation And Prediction Of Wind Speed With Multi-scale Process And Multivariable Constraint

Atmosphere reanalysis data will be interpolated from high resolution to low resolution according to the requirements of the application.Linear interpolation and exponential interpolation is usually used,of which the bias is big.Thus,the single-scale Gaussian process regression is usually used in 2-D space series interpolation,and wind values are usually calculated by recurrent neural network in 1-D time series prediction.However,these methods all base on the hypothesis that the evaluated values are single signals and are independent to other variables,which ignore the physical significance and physical constraints of meteorological data.Wind speed results from multiple physical processes,having a complex nonlinear relationship with temperature,pressure,density and other meteorological variables.Wind speed sequence is a signal with multi-scale process and multivariable restrictions in which there is lots of important hidden information.This thesis will improve the interpolation and extrapolation of wind speed sequence in different aspects to increase the accuracy.The thesis emphasized on several aspects as follows:(1)In this thesis the Gaussian Process Regression method for wind speed space interpolation is improved with a proposed multi-scale interpolation kernel function.Usually the kernel function consists of a single-scale covariance function and an independent Gaussian noise.Based on the characteristics of the multi-scale process of wind speed sequence,the new combined kernel function consists of a large-scale wind speed covariance function,a middle-scale and small-scale wind speed covariance function,a spatial correlated periodic covariance function,a spatial correlated noise covariance function and an independent Gaussian noise.The kernel function is modeled with interval of 2 degree and the target output is reanalysis data with interval of 1 degree.The results show that the interpolation prediction of the kernel function with multi-scale is effective.The RMSE of the approach in this thesis is smaller than the one of linear interpolation,exponential interpolation and the interpolation with single square exponential covariance kernel function.Better results have been achieved in zonal wind component and meridional wind component prediction.In interpolation calculation of wind speed,the error generated by the interpolation of zonal wind component and meridional wind component is much smaller than the one generated by the norm of wind vector and the angle of wind vector.(2)The data format in the prediction of wind speed time series is improved with the sequence elements changing from single zonal wind(or meridional wind)scalar to the wind state vector with multivariable constrains.The wind speed state vector made up by some of zonal wind(or meridional wind),geopotential height,relative humidity and temperature will be applied to the LSTM based Sequence to Sequence prediction approach.The reanalysis data from 1948 to 1972 are chosen as the training set and cross-validation set,and the reanalysis data from 1973 to 1975 are chosen as the testing set.Note that the prediction from state vector to state vector shows a good result,in which the state vector made by wind speed component,temperature and geopotential height generates the smallest RMSE,while the state vector containing relative humidity doesn't work well.In conclusion,if the elements of wind speed state vector have strong relationship with each other,the RMSE of prediction will be lower than the RMSE of prediction without wind state vector.Otherwise,the results of prediction will not be improved.