综合天气相似分析方法及其气象预报服务应用

A Comprehensive Weather Similarity Analysis Method and Its Application in Meteorological Forecast Services

摘要:

为改进传统“切片”式天气形势相似分析方法存在的不同切片相似结果不一致、预报稳定性欠佳问题，借鉴大数据思维，将天气系统视为一个由高中低层大气相互配合、静力热力动力条件相互影响的综合体，以多种气象要素再分析格点资料为基础，采用机器学习PCA方法对原始数据进行降维、浓缩，经归一化处理后构建出适于综合天气相似分析的样本衍生特征因子矩阵;然后使用KNN算法计算样本间各特征维度的相似距离、并结合方差贡献率赋予其相应的权重，最终按综合相似距离大小排序给出目标样本在历史天气形势库中的综合最相似序列，从而实现对传统相似天气预报方法的升级改进。对比分析和测试应用表明，该方法可提供多要素、多层次“立体”综合相似下的一致性结论，有助于预报员更好地理解天气系统结构和演变过程、进而更准确地研判可能发生的相关天气现象，在精细化气象预报服务方面有良好的应用前景。在2023年以来的几次广西区域性极端降水气象预报服务中，该方法取得了较为显著的应用效果。

Abstract:

This article introduces a novel method that draws on big data thinking, treating the weather system as a comprehensive entity in which the interactions of the high, middle, and low-level atmospheres, as well as the influences of static, thermal, and dynamic conditions, are considered. It utilises a novel approach to comprehensive similarity assessment through situation field analysis, using derived data from numerical weather models and reanalysed grid data of various meteorological elements as its fundamental characteristics. The approach begins by employing the machine learning Principal Component Analysis (PCA) method to condense the features of the original grid field data, making it adaptable to the resource processing capabilities of conventional business platforms. Subsequently, the derived dimensional feature data of different meteorological elements at various spatial levels are normalised to ensure a balanced effect when participating in similarity calculations. The constructed sample-derived feature factor matrix, suitable for comprehensive weather similarity analysis, undergoes calculation of the similarity distance for each feature dimension among the samples. Based on the variance contribution rates of the initial field information contained in the data from different “principal component” dimensions, different weights are assigned to the similarity distance results of each dimension, yielding a comprehensive similarity distance. Finally, using the K-Nearest Neighbours (KNN) algorithm, the method provides the most comprehensive similar sequence in the historical weather situation database for the target sample, thus upgrading and improving traditional methods of similar weather forecasting. This method provides a multi-element and multi-level “stereoscopic” comprehensive similarity, aiding forecasters in better understanding the structure and evolution of weather systems and, consequently, more accurately assessing the possible occurrence of related weather phenomena. Comparative analysis and testing applications indicate that the results of comprehensive similarity analysis are superior to traditional “sliced” similarity analysis, which only targets single meteorological elements or altitude levels, particularly in terms of matching critical weather system positions and strength features. It resolves issues such as inconsistent results of similar weather situation analysis for different “slices” and poor forecast stability. This method provides more direct and efficient assistance in weather analysis and forecasting and holds promising prospects for refined meteorological forecasting services. In several instances of extreme precipitation meteorological forecasting services in the Guangxi region since 2023, this method achieves significant application effectiveness.