Forecast Inconsistency Of Medium-range Numerical Weather Prediction Based On TIGGE Datasets

Based on the 500hPa geopotential height, the 850hPa temperature and the mean sea level pressure forecasts from the ECMWF, NCEP and CMA in the TIGGE datasets, the characteristics of the forecast inconsistency for the control and ensemble-mean forecasts and the comparison of their characteristics have been conducted by using Jumpiness index and other different forecast jumps-the "flip", "flip-flop", "flip-flop-flip" and so on. The influence of the ensemble prediction method on inconsistency and the methods to reducing the forecast inconsistency are also discussed. The conclusions are as follows:Firstly, in terms of the period-average inconsistency features, all average the period-average inconsistency indices increase with the forecast range in agreement with the practical experience that the forecasts are usually more consistent at short forecast ranges. And for ensemble prediction system, the ensemble-mean forecast is less inconsistent than its corresponding control forecast, especially at long forecast ranges, which indicates that the forecast inconsistency could be reduced using the ensemble prediction method. And both for the control forecast and ensemble-mean forecast, the forecast consistency of ECMWF is better.Secondly, in frequency statistics of the inconsistency, the frequencies of the "flip", "flip-flop" and "flip-flop-flip" are in descending order. For these three types of forecast jumps, the frequency of ensemble-mean forecast is significantly lower than that of the control forecast especially at long forecast ranges. This indicates that the ensemble-mean forecast is more consistent than its corresponding control forecast, which also shows that the forecast inconsistency could be reduced using the ensemble prediction method. The frequency variation of parallel "flip", parallel "flip-flop" and parallel "flip-flop-flip" indicates that the control forecast and ensemble-mean forecast have large difference at long forecast ranges. And the correlation coefficient of their Jumpiness indices also confirms this conclusion.Thirdly, the sensitivity of the forecast inconsistency shows that the period-average inconsistency has a strong sensitivity to the area and parameter. And the sensitivity of the control forecast to the area and parameter is stronger than that of ensemble-mean forecast. The smaller the studied area, the larger the period-average inconsistency becomes, which indicates that the inconsistency intensity is stronger. As the weather climate of the selected areas is not the same, the period-average inconsistency is different. For different variables, the period-average inconsistency is also various. The period-average inconsistency index of mean sea level pressure is the maximum, the result of 500hPa geopotential height is the second, and the minimum result is the 850hPa temperature. That is to say, the inconsistency intensity of temperature is lower than the geopotential height results. And the frequency of different forecast jumps and the difference of the inconsistency between control forecast and ensemble-mean forecast show little sensitivity to the choice of the area, time and parameter.At last, in terms of the improvement in forecast inconsistency, the intensity and frequency of the forecast inconsistency are improved to some extent by utilizing the methods of multiple linear regression and bias-removed mean with multi-lead time forecasts, especially significantly at long forecast ranges. But at short forecast ranges, the correction results of these two methods are not ideal. These may be related with the theory of these methods and some inherent features of models and need to be further analyzed. For multi-model ensemble forecast methods, the intensity and frequency of the forecast inconsistency are significantly improved by using the simple multi-model ensemble mean method, which is related with its own theory. However, for the superensemble forecast method, the improvement effects are limited and close to the results of multiple linear regression and bias-removed mean with multi-lead time forecasts, which may be because that its theory is similar to that of the multiple linear regression and bias-removed mean with multi-lead time forecasts.