基于机器学习分类算法的臭氧浓度等级预报在长沙的应用

为准确预报臭氧浓度等级，基于EC＿THIN全球天气模式产品和我国自主研发的CMA＿GFS全球天气数值预报产品以及华南GRACEs大气成分模式输出产品，融合气象和环境观测数据，使用6种机器学习智能算法，构建耦合数值预报模式和机器学习的混合模型，旨在充分发挥数值预报与机器学习智能算法的优势和互补协同作用，实现臭氧浓度等级预报准确度的跨越式提升。共设置4个控制试验，选取不同的特征产品，依次使用机器学习经典分类算法对长沙市未来4天的臭氧浓度等级进行分类预报，取测试准确度最高的模型输出结果作为结果统计。发现：最优模型1～4天的测试准确度分别为81.7%、81.7%、78.3%、60.9%，大大高于大气成分模式预报和预报员经验，达到预期设计目标；高质量的天气模式产品对模型贡献大，而大气成分模式产品对模型贡献有限；模型3天以内预测性能较好，低等级预测性能较好，高等级预测性能一般。提出解决方案供讨论：增加高等级样本数量，增强模型对此类事件的识别能力；加强高等级臭氧污染的机理分析，组合出更精炼的因子供模型使用。

PREDICTION PRACTICE OF OZONE CONCENTRATION LEVEL BASED ON MACHINE LEARNING CLASSIFICATION ALGORITHM

In order to accurately predict the ozone concentration level, this paper used the EC_ Thin global weather model products, global numerical weather prediction products from GRAPES_GFS, which are the graphs independently developed in China, and South China GRACEs atmospheric composition model output products to integrate meteorological and environmental observation data, and used six machine learning intelligent algorithms to build a hybrid model that coupled numerical prediction models with machine learning, in order to give full play to the advantages, complementarity and synergy of numerical prediction and machine learning intelligent algorithms, and improve the accuracy of ozone concentration level prediction by leaps and bounds. Four control experiments were set up for which different characteristic products were selected and the classical classification algorithm of machine learning was used to classify and predict the ozone concentration level in Changsha for any four days to come. The model output with the highest test accuracy was taken as the result statistics. It is found that the test accuracy of the optimal model in 1 to 4 days is 81.7%, 81.7%, 78.3% and 60.9% respectively, which is much higher than the atmospheric composition model prediction and forecaster experience, and achieves the expected design goal. The contribution of high-quality weather model products to product quality is large while that of atmospheric composition model products is limited. The prediction performance of the model within 3 days is good, and low-level prediction performance is good, and high-level prediction performance is general. The following solutions for discussion are proposed: the number of high-level samples is to be increased; the recognition ability of the model is to be enhanced for such events; the mechanism analysis of high-grade ozone pollution is to be strengthened; and more refined factors for the use of the model are to be combined.