A Theoretical And Methodology Research Of GRAPES-global Improvement In Numerical Weather Prediction Using The Past Data

The initial value error and the imperfect numerical model are usually considered as the error sources of numerical weather prediction (NWP). However, model errors can be inversely estimated by using the past multi-time observations and model output and then removed to improve model performance. Two essential questions should be answered for improving NWP by using past data:(1) how to estimate the past model error and (2) how to extrapolate the model error to the future. According to this two questions, the following studies present as:(1)A theoretical method is presented and assessed firstly, which proposed trapezoidal approximation to obtain past model error and Lagrange polynomial function to extrapolate. As assumed that the past data and past model error are exactly accurate, the order of Lagrange polynomial and the times of past data must stay the same. However, the errors are inevitably exist in past data and past model error estimated by trapezoidal approximation, so the theoretical method is impracticable in operational NWP model because the high order Lagrange polynomial is sharply sensitive to the error.(2) On the issue of estimating the past model error, a iteration method has been occupied in this study. The convergence of this method has been confirmed with ideal experiments and series of real tests. Compared with systematic error (5year averaged) of GRAPES, it was confirmed as well that the model error obtained by iteration brings into corresponding with GRAPES forecast error.(3) A new method for improving NWP by using past data has been presented, which combined least square method with the inverse problem. The advantage of this method is that it is more practical and targeted, because error functions for extrapolating can optionally utilized according to different component of model error and not limited by the times of past data.(4) Extrapolated by the past model error,0-order of the new method has been applied in GRAPES and two series of real tests are presented. In terms of mean circulation field, mean error of500hPa and850hPa and zonal averaged error profile, the new method sharply reduced aspects of GRAPES forecast error. The averaged scores (bias, rmse and ACC) also displays that the new method can remove the model error and improve GRAPES forecast.In summery, a new method for improving NWP by using past data was developed, which can practically applied in operational NWP model without developing specific model or modifying pre-exit model intensively. Combined least square method with the inverse problem, the new method can use error functions for extrapolating and past data more optionally. Series of tests has confirmed the validity and effectiveness.