The Effects Of Background Error Covariance Simulation On Data Assimilation And Forecast

Background error covariance determines the weight between the observation and the background, and plays a crucial role in qualifying in the analysis of data assimilation. For a better understanding and probing into the impact of different background samples to background error covariance modeling, and to data assimilation and numerical forecasting, this paper will focus on the samples modeling problem of the background error covariance, four sets of background error sample are generated by sample error simulation techniques, such as, Randomized Control Variables(RandomCV), Breeding of Growing Modes(BGM), grid-point random multivariate perturbations of the entire domain(MVP) and the NMC method (or the known "NCEP" method), and then background error covariance matrixes (B) are generated accordingly on the basis of the Weather Research Forecast Data Assimilation System(WRFDA). After the characteristic analysis of background error covariance, the pseudo single observation tests, the data assimilation cycles and forecast, the characteristic of the background error covariance and their effects to data assimilation and numerical forecast are discussed in detail, the results show that:(1) The results of background error covariance structure characteristic analysis show that:the background error covariance of NMC and RandomCV, who’s eigenvalues of the first model are larger than BGM and MVP, which means their error amplitudes are more significant, also means their weight of observation will be increased during data assimilation; The horizontal lenthscale implies the influence area of the observation information, while the lenthscale of RandomCV and BGM are basically larger than the rest.(2) The conclusion drawn from the temperature pseudo single observation tests, which show that the tests’ results of each method fitting well with their corresponding background error covariance matrix characteristics. And the influence area of RandomCV method and NMC are basically broader than other methods, for their higher lengthscale values, meanwhile BGM and RandomCV show larger analytical increment for their larger eigenvalues.(3) Results of the data assimilation cycles and numerical forecast show that:values of the root mean square error (RMSE) and absolute bias(Abias) of NMC and RandomCV are lower during the implementation of their background error covariance matrixes in data assimilation cycles and numerical forecast, while the overall effect of the background error covariance matrix contain in the WRF model is not so good as the two methods mentioned above, MVP’s is the worst still.(4) Before the samples for generating the background error covariance by NMC method are gained, what we need not only the large amount of reanalysis data, but also the numerical modelling at least a month. The RandomCV method used in this paper,who’s ensemble members are calculated under the perturbation of the initial condition by the background error covariance matrix contain in the WRF model, which is the background field samples for calculating the background error covariance, that achieves almost the same effect as the NMC in data assimilation and forecast, but a lower computational and manual cost.