Research On The Interpretation Of Numerical Forecast Productions For Day-by-day Precipitation Forecast

With the artificial neural network, a new research is carried out on the interpretation of 48h numerical forecast productions of Japan Fine Grid Precipitation Model and T213 Model on May&June in 2002~2003. Firstly, three regions are set as basic forecast regions by the use of Cluster Analysis according to the day-by-day precipitation datasets for Guangxi 89 stations on May&June 1951~2000. Three schemes are designed to make the operational prediction experiment on May and June 2004.Scheme 1: according to operational rule, the average precipitation of 3 regions is devided into 5 leveles (rainless, light rain, middle rain, heavy rain and storm rainfall). Taking the 24h average rainfall level as the forecast object, choosing elmetary forecast factors from T213 model and Japanese numerical weather forecast productions base on correlation. The artificial neural network rainfall level qualitative forecast model with small network construction is set up using the components of elemetory foctors dealed with Priciple Component Analysis (PCA). And actual operational experiment is made by the use of the neural model. The capability of the neural model is analysised, and the improved measure is presented.Scheme 2: depending on the improved measure in sheme 1, a new test is made for the same rainfall level as the scheme 1. To keep the useful information of the bst factor, the Japanese model factor, the PCA is only exerted on the T213 model factors. The new neural model is set up with the components factors from PCA exerted on T213 and the Japanese model factor. The result suggests the improved method is effective. The TS evaluation of 3 regional neural models is 0.55, 0.5 and 0.26 respectively for middle rainfall, and they are better than T213 and Japanese model, can be used as referenced predctions.Scheme 3: Base on the successful qualitative precitation forecast test, the quantitative precipitation forecast test is made. The neural network models for the 24h average rainfall of 3 areas are set up making the use of the same method as scheme 2 to do the actual operational prediction experiment. The statisctical result for the experiment shows that the neural model is better than T213 on the mean absolute error,the maximum error and the credible forecast percentage, its predicted capability for middle rainfall is superior to the T213 model in the corresponding period. The future for this sheme is attractive.Analysising the results of 3 schemes, the method presented in the article is successful to build a neural network model with small network struction and rich forecast information by keeping valuable forecast factor and using Principle Component Analysis on the others numerous factors. It is a new test to utilize fully a large amout of the numerical forecast prodution. It's an effective method to improve the operational forecast level.