GRAPES模式温度产品误差统计订正方法比较研究

Comparison on Statistical Correction Methods of Temperature Product Error in GRAPES Model

为了提高GRAPES模式下的温度产品的准确性,本文采用2021年全年GRAPES模式下的温度数据作为预报数据,欧洲中心的ERA产品作为实况数据,数据点分布为格点分布。研究区域为30.61˚N~30.73˚N,103.78˚E~103.90˚E,将该范围内的25个网格点上的数据,按照每1个小时一次的时间间隔,通过Kriging插值处理预报数据的初值场插值到靠近温江站的格点30.7˚N,103.8˚E,得到较为稳定的预报数据后,运用线性拟合从总体距平值最小的角度分析最为符合预报–实况温度的拟合线;运用三次样条插值拟合从点与点之间的斜率关系分析最为符合预报和实况温度的拟合线;还运用神经网络分析训练,以尽量减小预报温度和实况温度间的异常大的差为目的,来拟合最符合预报和实况温度的曲线。对于订正结果,分析对比订正后拟合结果与实际观测数据的残差,来判定某一订正方法的可信度高低。结果表明:1) 经过Kriging插值处理过后的初值场,残差数值大小从整体上已经呈现出小值多,大值少,和实况数据较为贴近。2) 三种订正方法的拟合结果,均比原来Kriging插值的初值场更加贴近实况数据,残差数值小的区域均数值点更加密集。3) 三种订正方法对比,三次样条插值拟合的残差数值大的区域有较多的拟合数据点,尤其是残差超过5℃的区域,容易引起部分时间点,拟合和实况有较大出入;线性拟合与神经网络分析训练的拟合结果的残差数值小的区域数据分布均很稀疏,但是二者对比,神经网络分析训练的拟合结果,拟合数据与实况数据的残差分布大半都集中在−2℃~2℃的小值区域,结果与实况大致相同。

In order to improve the accuracy of the temperature products under the GRAPES model, this paper uses the temperature data under the GRAPES model for the whole year of 2021 as the forecast data, the ERA product of the European Center as the live data, and the data points are distributed as grid points. The study area is 30.61˚N~30.73˚N, 103.78˚E~103.90˚E. The data on the 25 grid points within the range are processed by Kriging interpolation according to the time interval of every 1 hour. The initial value field interpolation of the forecast data to the grid point 30.7˚N, 103.8˚E near Wenjiang Station, after obtaining relatively stable forecast data, use linear fitting to analyze the fitting line that best fits the forecast and actual temperature from the angle with the smallest overall anomaly value;use cubic spline interpolation. The fitting is based on the slope relationship analysis between the points and the fitting line that best fits the forecast and actual temperature;a neural network is also used to analyze the training, in order to minimize the abnormally large difference between the forecast temperature and the actual temperature, to fit Combine the curve that best matches the forecast and actual temperature. For the correction results, the residuals between the fitting results after correction and the actual observation data are analyzed and compared to determine the reliability of a correction method. The results show that: 1) In the initial value field after Kriging interpolation, the overall residual value has more small values and less large values, which is closer to the real data. 2) The fitting results of the three correction methods are all closer to the real data than the original Kriging interpolation initial value field, and the mean value points in the area with small residual values are denser. 3) Compared with the three correction methods, the area with large residual value of cubic spline interpolation fitting has more fitting data points, especially the area with residual error exceeding 5˚C, which is easy to cause some time points, fitting and there is a big discrepancy between the actual situation;the data distribution in the area where the residual value of the fitting result of the linear fitting and the neural network analysis training is small is very sparse, but the comparison between the two, the fitting result of the neural network analysis training, the fitting data and the actual data are very sparse. Most of the residual distribution of the data is concentrated in the small value region of −2˚C to 2˚C. The results are much the same as the actual situation.