| With the advances of observational techniques and numerical prediction models, the simulation skill of numerical weather prediction model continues to increase. It is known that the numerical prediction models only approximate the real atmosphere. Moreover, the inaccurate initial and boundary conditions of the model and chaotic characteristics of the atmosphere itself can also cause the error between the numerical model forecasts and the real atmosphere. Therefore, Error correction has become one of the effective ways to improve the accuracy of numerical prediction models.In this paper, the spatial and temporal distribution characteristics of forecast error is analyzed by Empirical Orthogonal Function (EOF) and Empirical Mode Decomposition (EMD) method with T213500hPa geopotential height data from2003to2007. Based on the above analysis of forecast error, non-systematic forecast error is estimated by using BP neural network to establish its prediction model from two aspects. One error prediction model is established based on the time-dependencies and nonlinear variation over time of the numerical model forecast error. The other error prediction model is established by finding the mapping relationship between the state variables and the prediction error. The main conclusions are as follows:(1) The space distribution of systematic forecast error is displayed by calculating the average forecast error over time at all grid points. The results show that the forecast error over the whole area is negative deviation. Its absolute value takes on a pattern of high in the West and lower in the East.(2) Standard deviation of forecast errors is distributed homogeneously with longitude generally. From low latitude to high latitude, it increases, and then decreases with latitudes, expect some individual areas, such as Qinghai-Tibet Plateau, the Sichuan Basin, and south of Qinling area.(3) It is found that the forecast errors do not meet the normal distribution seen from QQ and frequency histograms on forecast error. EMD decomposition results show the variation of forecast error with time consists of the shock component of multiple time scales. The shock cycle and magnitude at each time scale varies with time. The variation of forecast errors with time has non-stationary characteristics. However, it can be approximated as a stationary process for the small variation of the trend term from EMD decomposition.(4) Based on the time-dependency, the three-layer BP artificial neural network is employed to establish the error predication model to estimate non-systematic forecast errors in the future time. Its effectiveness is tested with1732forecast samples from T213model. The results show that the established predication model has a good ability for estimating the non-systematic forecast errors in the future time. The characteristic of estimated on-systematic forecast error is consistent with the truth. Compared with the systematic error correction, the error correction from the error prediction model based on BP network can further improve the forecast skill of numerical prediction model.(5) The other error prediction model is established by using BP ANN to obtain the mapping relationship from state variables to forecast error, and its validity is tested with the same1732samples. The preliminary experimental results show the effect of this error prediction model is unstable. Compared the systematic error correction, the error prediction model based on BP network can improve the forecast skill of the samples that the ration of RMSE error between corrected forecast and non-corrected forecast is less than0.5. The Percentage of valid correction increases from1%to34%. However, the error prediction method based on BP network method slightly increases invalid corrections, and its percentage increases from39%to44%. |
PDF Full Text Request