基于全场信息的数值预报产品释用方法研究

对于短期气候时间尺度预报来说, 短期内尚无有效方法提高数值预报产品精度, 建立一种能综合有效利用全场信息的非线性释用方法, 通过统计释用提高短期气候的预报准确率, 是一条可行的途径。通过CCA-BP法建立的典型因子, 可以代表因子场与站点预报要素之间的大部分协方差关系, 使因子与站点要素相关性大为提高, 进而通过神经网络技术 (BPNN) 建立非线性预报模型, 实现上述目标。以平潭、福州站10月旬平均温度、降水距平百分率预报为例, 分别使用CCA-BP-BPNN模型和插值模型对1983-2001年资料建立预报方程, 对2002-2005年的试报结果表明:解释预报对数值预报产品做了较大修正,使预报产品具有一定的使用意义及参考价值。从各项评价指标来看,CCA-BP-BPNN模型优于插值模型。该方法为提高短期气候数值预报产品的释用精度提供了一个值得参考的途径。

An Interpretation Scheme on Numerical Products of Prediction System Based upon Full-scale Information

It is difficult to improve the accuracy of short-term numerical climate prediction in the scale of a month or so at present. A nonlinear interpretation model that uses full-scale information of the numerical products is proved to be effective in raising the accuracy of shot-term climate prediction. Besides, the interpretation method is the only way to interpret the grid data of numerical system to stations. Nowadays, most interpretive approaches are based on the MOS (Model Output Statistics). Some potential imperfectness of this method lay in the facts that it can't use full-scale information synthetically and effectively, besides some systematic errors. The change of mean meteorological element may not be reflected from the grid data around the stations, however, it may have some uncertain connections with other grid data beyond the station. If the factors of two dimensions which reflect full-scale information are used, the result of prediction may be improved. Canonical variables are constituted using CCA-BP method, which mainly represents covariance between factors and the predictive variables. The relativity between new-formed composed factors and the predictive variables increases significantly, hence a nonlinear regression model is developed by BPNN method. However, the canonical variables may be influenced by some local change greatly. Thus the canonical variables in the predictive equation should be examined through long-term stability test, to use those stable and effective ones. The interpretation scheme is used to establish the predictive equations for 10-day average temperature and rainfall anomalous percentage in October of Pingtan and Fuzhou, China. The 4-year forecast experiment results indicate that interpretation improves the numerical prediction products remarkably and the prediction accuracy increases evidently. According to the evaluating index, the CCA-BP-BPNN (abridged as C-B method below) is superior to the interpolation method and gives a new way of interpretation prediction of the short-term climate prediction products. Statistics of 4-year prediction results using the present C-B method from 2002-2005 shows, both the prediction accuracy (TT) and anomalous relationship coefficients (ACC) of temperature and precipitation of Pingtan and Fuzhou increase comparing with the interpolation method results and the model outputs of MM5, while the mean absolute errors (MAE) and mean square-root errors (MSE) decrease.The 23 years long dataset used in the experiment makes the monthly prediction variables a sample size of 23 and the 10-day prediction variables a sample size of 69. It is statistically acceptable but not as good as expected. For the interpretation prediction, longer the data series brings better results. Besides, selection and composition of the factors are endless jobs for regression model establishment. With the accumulation of numerical prediction products and the factor optimization procedure used in the CCA-BP-BPNN model improvement, the forecast accuracy will be increased in the future. Limited by the shortage of data and capability of computing, only 2 southeast coastal stations are involved in the test, but the results are satisfactory. However, the adaptability of this model in northern and western region with great climatic variability still needs validating with a mass of case studies.