TEMPERATURE FORECAST CORRECTION METHOD AND EVALUATION BASED ON RADIAL BASIS FUNCTION NEURAL NETWORK
-
摘要: 根据中央气象台自2017年10月—2018年9月20:00起报未来72 h 0.05 °×0.05 °分辨率格点日最高、最低温度指导预报和国家气象信息中心格点温度实况,应用Matlab神经网络工具箱提供的newrbe函数,建立基于径向基函数(RBF)神经网络的温度预报模型,对2018年10月—2019年9月RBF预报产品进行格点检验评估,并与同期的EC模式预报产品做了对比。结果表明:(1)通过RBF模型订正后的24 h、48 h和72 h日最高和最低温度预报准确率较中央气象台指导预报(NMC)分别提高了7.21%、6.98%、5.48%和5.67%、4.46%、4.47%,均为正技巧,且春、夏、秋季预报订正效果要好于冬季;(2)分区域预报检验来看,除海源、同心、彭阳的最高温度预报和海源、惠农的最低温度预报误差偏较大外,其他区域的误差基本都小于2 ℃。特别是对强降温、霜冻天气的温度预报准确率高于NMC,对预报员有一定的参考价值。Abstract: The present study uses the daily maximum and minimum temperature 72 hours guidance forecast with 0.05°×0.05° grid resolution released by the National Meteorological Center (NMC) for the period from October 2017 to September 2018, the actual temperature observed by the national meteorological information grid, and the newrbe function provided by the Matlab neural network toolbox to establish a temperature forecast model based on the radial basis function (RBF) neural network. Then, the temperature forecast model is used to conduct grid inspection and evaluation of the RBF forecast products for the period from October 2018 to September 2019. It is also compared with the EC model forecast products for the same period. The results show that: (1) The accuracy of the daily maximum and minimum temperature forecasts at 24h, 48h and 72h after the correction of the RBF model is increased by 7.21% and 6.98%, 5.48% and 5.67%, 4.46% and 4.47%, respectively. They are all positive skills, and the correction in spring, summer and autumn forecast is larger than that in winter. (2) According to the subregional forecast inspection, the errors are basically less than 2 degrees, except for those of the maximum temperature of Haiyuan, Tongxin, and Pengyang and the minimum temperature of Haiyuan and Huinong.In particular, the accuracy of temperature forecast for severe cooling and frost weather is higher than that by the NMC, and the former can serve as reference for weather forecasters.
-
Key words:
- radial basis /
- neural network /
- maximum temperature /
- minimum temperature
-
[1] 薛志磊, 张书余. 温度预报方法研究及其应用进综述[J]. 干旱气象, 2012, 30(3): 459-464. [2] 陈豫英, 陈晓光, 马金仁, 等. 基于MM5模式的精细化MOS温度预报[J]. 干旱气象, 2005, 23(4): 52-56. [3] 陈豫英, 陈楠, 王素艳, 等. MOS方法在动力延伸期候平均温度预报中的应用[J]. 应用气象学报, 2011, 22(1): 86-95. [4] 吴启树, 韩美, 郭弘, 等. MOS温度预报中最优训练期方案[J]. 应用气象学报, 2016, 27(4): 426-434. [5] 李佰平, 智协飞. ECMWF模式地面温度预报的四种误差订正方法的比较研究[J]. 气象, 2012, 38(8): 897-902. [6] 罗聪, 曾沁, 高亭亭, 等. 精细化逐时滚动温度预报方法及检验[J]. 热带气象学报, 2012, 28(4): 552-556. [7] 张成军, 雷学锋, 李娜, 等. 基于动态最优PP法的乡镇温度预报技术[J]. 干旱气象, 2019, 37(3): 508-514. [8] 马学款, 普布次仁, 唐叔乙, 等. 人工神经网络在西藏中短期温度预报中的应用[J]. 高原气象, 2007, 26(3): 491-495. [9] 吴君, 裴洪芹, 石莹, 等. 基于数值预报产品的地面温度BP-MOS预报方法[J]. 气象科学, 2017, 27(4): 430-435. [10] 张天虎, 鲍艳松, 钱芝颖, 等. 基于BP神经网络与遗传算法反演大气温湿廓线[J]. 热带气象学报, 2020, 36(1): 97-107. [11] 雷彦森, 蔡晓军, 王文, 等. 遗传算法优化的BP神经网络在地面温度多模式集成预报的应用研究[J]. 气象科学, 2018, 38(6): 806-814. [12] 郭庆春, 李力, 张冉, 等. 我国气温变化的神经网络预测模型[J]. 热带气象学报, 2009, 25(4): 483-487. [13] 农吉夫. 基于MATLAB的主成分RBF神经网络降水预报模型[J]. 热带气象学报, 2008, 24(6): 713-717. [14] 高领花, 苏炳凯, 汤剑平. RBF神经网络集成多个气象数值产品的方法研究[J]. 浙江大学学报(理学版), 2006, 33(2): 223-230. [15] 熊聪聪, 潘璇, 赵奇, 等. 多模式集成的RBF神经网络天气预报[J]. 天津科技大学学报, 2014, 29(1): 75-78. [16] ZHU M, PANG L X, XIAO Z J, et al. Temperature Drift Compensation for High-G MEMS Accelerometer Based on RBF NN Improved Method[J]. Applied Sciences, 2019, 9, 695. [17] 陈广. 基于改进PSO_RBF神经网络的气温预测模型[D]. 兰州: 兰州大学, 2015. [18] LI J Y, ZHAO Z D, LI Y S, etc. Short-term PV/T module temperature prediction based on PCA-RBF neural network[J]. Earth and Environmental Science, 2018, 121: 052045. [19] 任巧丽, 马乾. PCA−RBF神经网络对兰州市日平均气温的预测[J]. 宜宾学院学报, 2018, 18(12): 111-115. [20] 张成军, 纪晓玲, 马金仁, 等. 多种数值预报及其释用产品在宁夏天气预报业务中的检验评估[J]. 干旱气象, 2017, 35(1): 148-156. [21] 陈迪, 陈豫英, 马金仁, 等. 不同时间尺度的MOS方法对宁夏温度预报的影响[J]. 干旱区地理, 2019, 42(1): 94-102.