Research Of Model Output Statistics Technique And Its Application In Short-term And Medium-term Local Weather Forecast

Local weather forecast has always always been concerned by weather forecaster. Because of the lower resolution global forecast model, near-surface boundary layer physical processes, as well as dynamic processes and thermal processes by the impact of underlying surface, simulation results are not satisfactory at the local scale. So the global forecasting models can not produce the satisfactory meteorological elements forecasting result. To obtain more accurate forecasting results, the results from global forecast model are downscaled. Because the dynamic downscaling need more resources and better conditions for the computer, the statistical-dynamic downscaling methods (ie, the model output statistical techniques) just can make up for lack of dynamic downscaling, It is less to calculation and easy to operate, very suitable for grass-roots meteorological stations. In many developed countries, it was the backbone of modern weather forecasting, especially in the numerical model of the post-processing, it was an indispensable part. In this paper, based on previous studies and actual needs, some problems about model output statistics are need to research comprehensively from two aspects of research and practical application.Then, a complete solution scheme is presented.The model output statistics (MOS) technology is core idea of the concept in full-text. According to research needs, the predictor selection and prediction modeling in model output statistics are study depthly. Then many encouraging results have been achieved. Study stations include 25 weather stations in Qinghai Province and the 23 weather stations around the Bohai region. Test data include daily model output products of National Meteorological Center T213L31 the medium-term global numerical prediction as well as the above-mentioned weather site observations data. Datas is from January 1,2003 to December 31,2009. Maximum temperature, minimum temperature and precipitation are as trial weather object.The stepwise regression methods associated with correlation coefficient are as predictor selection of statistical downscaling. The experiments have shown that: Although the number of preparation factors on the overall level is less sensitive to the forecasting hitrate, but the number of different preparation factors will have an impact for forecast results in a few sites.From F value test can be seen that the forecasting result is less sensitive to F values, but the F value can not be smaller, because the smaller F value will cause the forecast results unstable. When a historical observation is as a predictor, the forecast results, especially short-term (24,48 hours) forecast TS have been significantly improved in winter. For the summer, the effect is not very obvious.In temperature modeling represented by continuous element, neural network are in-depth study. Results show that: Random weights which have a Gaussian distribution with zero mean and a standard deviation of one and multiple weights initialization and dynamic number of neurons in neural networks (ANN) can give a better approximation of the true value. Compared with multiple line regression (MLR) method, In some stations, TS from ANN is slightly higher in winter minimum temperatures and summer maximum temperature forecast. But ANN have the same forecasting level with MLR. It shows that the generalization ability of network is not strong, more research is needed.For precipitation forecast, the phenomenon of false alarm from numerical prediction results is overcome by improved k-NN method to a certain extent. Neural networks 2 (ANN2) use TS of training samples associated with TS of validating samples as the objective rules. Neural networks 3 (ANN3) only use the TS of training samples as the objective rules. Optimal weights model produced by ANN2 has stronger generalization ability than model builded by ANN3. Forecasting results from ANN2 are better than the result predicted by ANN3. In generally, model builded by ANN3 exit instability and generalization capacity problem. The forecasting effect of k-NN method is fairly stable, although the overall TS of k-NN prediction is lower than TS of ANN2, k-NN prediction results still available. Both methods have advantages, so in real operations, you can use both methods, learn from each other. Purpose of the study is to serve the needs of practical application. In this paper, there are two examples, short-term numerical prediction interpretation system in Qinghai province and the medium-term numerical prediction interpretation system in Huludao set up by us, the framework, advantages and disadvantages of the two systems will be introduced in terms of the application aspects.Short-term numerical prediction interpretation system in Qinghai province is our first set of self-developed system. The system has been put into operational in Qinghai Bureau of Meteorology.24-hour temperature forecasts are given a higher rating that has a higher accuracy rate. Medium-term numerical prediction interpretation system in Huludao is comprehensively upgrade, which lenan advanced concepts and techniques from short-term numerical prediction interpretation system in Qinghai province. The framework of the whole system are re-designed, So the entire system can better adapt to the various research and business needs.