| 降水统计预报模型的模拟性能分析北大核心CSCD |
| |
| 引用本文： | 陈以祺,吴香华,刘鹏,刘端阳.降水统计预报模型的模拟性能分析北大核心CSCD[J].气候与环境研究,2022,27(5):578-590. |
| |
| 作者姓名： | 陈以祺 吴香华 刘鹏 刘端阳 |
| |
| 作者单位： | 1.南京信息工程大学数学与统计学院，南京 2100442.南京信息工程大学大气科学学院，南京 2100443.南京交通气象研究所，南京 210008 |
| |
| 基金项目： | 国家重点研发计划项目2018YFC1507905，国家自然科学基金项目42075068、41505118、41975087、41605045 |
| |
| 摘 要： | 降水预报模型的性能与诸多因素有关,除了与研究区域特征和研究数据有关,还受到模型自身算法、统计模拟方法、性能度量指标等的影响。本文基于2015~2019年我国黑龙江省28个站点逐日降水、平均气温和平均相对湿度等地面常规气象资料,运用留出法、自助法等蒙特卡洛统计模拟和机器学习方法,首次系统研究了黑龙江省夏季逐日降水预报模型的性能和模型性能的空间分布特征。结果表明,对研究区域整体来说,BP(Back Propagation)神经网络和支持向量机的总体预报性能没有显著差异,ROC(Receiver Operating Characteristic)曲线面积值均高于76%,显著优于决策树。自助法估计的模型预报性能始终优于留出法,并且有助于提高评估结果的保真性。对研究区域单个站点来说,除个别站点以外,支持向量机的准确率和ROC曲线面积值均高于80%,并且呈现东南大西北小的空间分布趋势,该趋势与降水频率的分布基本一致。支持向量机在小兴安岭和张广才岭的总体预报效果较好,三江平原次之,松嫩平原较差;而敏感度在山区大,平原区小,中部和南部大,东部次之,西部和北部小;特异度空间分布则恰好与敏感度相反。
|
| 关 键 词： | 夏季降水预报 机器学习 留出法 交叉验证 自助法 |
| 收稿时间： | 2021-03-20 |
Analysis of the Simulation Performances of Precipitation Statistical Forecasting Models |
| |
| Yiqi CHEN,Xianghua WU,Peng LIU,Duanyang LIU.Analysis of the Simulation Performances of Precipitation Statistical Forecasting Models[J].Climatic and Environmental Research,2022,27(5):578-590. |
| |
| Authors: | Yiqi CHEN Xianghua WU Peng LIU Duanyang LIU |
| |
| Institution: | 1.School of Mathematics and Statistics, Nanjing University of Information Science and Technology, Nanjing 2100442.School of Atmospheric Sciences, Nanjing University of Information Science and Technology, Nanjing 2100443.Nanjing Institute of Transportation Meteorology, Nanjing 210008 |
| |
| Abstract: | The performance of a precipitation forecast model is related to many factors. In addition to research areas and research data characteristics, it is also affected by the model's algorithm, statistical simulation methods, and performance metrics. This paper is based on the daily rainfall, average temperature, and average relative humidity of 28 stations in Heilongjiang Province in China from 2015 to 2019, using Monte Carlo statistical simulation methods such as Hold-out, Bootstrap, and machine learning methods. For the first time, this paper systematically studied the performances of daily precipitation forecast models in Heilongjiang Province in the summer and the spatial distribution characteristics of the model performances. The results show that for the entire study area, the overall prediction performance of a BP (Back Propagation) neural network and support vector machine is not significantly different, and the value of the area under ROC cuvre is higher than 76%, which is significantly better than that of the decision tree. The prediction performance of the model estimated by Bootstrap is always better than that of Hold-out, and it helps improve the fidelity of the evaluation results. For a single station in the study area, except for certain stations, the value of accuracy and the area under ROC cuvre of the support vector machine are higher than 80%, and the spatial distribution trend is larger in the southeast and smaller in the northwest. This trend is basically consistent with the distribution of precipitation frequency. The overall prediction effect of the SVM (Support Vector Machine) model is better in the Xiaokingan and Zhangguangcai Mountains, followed by the Sanjiang and Songnen Plains. The sensitivity is higher in mountainous areas than in plain areas. The central and southern regions are larger, followed by the eastern region and then the western and northern regions. The spatial distribution of specificity is simply the opposite of that of sensitivity. |
| |
| Keywords: | |
| 本文献已被 维普 等数据库收录！ |
| 点击此处可从《气候与环境研究》浏览原始摘要信息 |
| 点击此处可从《气候与环境研究》下载免费的PDF全文 |
|
