机器学习在ENSO预测会商中的应用  

The application of machine learning in ENSO prediction consultation

作　　者：李晨彤 LI Chentong(National Marine Environmental Forecasting Center,Beijing 100081 China)

机构地区：[1]国家海洋环境预报中心,北京100081

出　　处：《海洋预报》2022年第1期91-103,共13页Marine Forecasts

基　　金：国家自然科学基金青年科学基金(41907180)。

摘　　要：基于多模式集合预报的思想,利用可解释机器学习方法——决策树算法建立了多模式ENSO预测结果智能会商系统。分别使用基于Boosting的GBDT、XGBoost、lightGBM和基于Bagging的RF 4种决策树模型方法,结合随机搜索交叉验证、网格搜索交叉验证两种超参数调整方法对决策树模型的超参数进行优化调整,根据不同超前预报时效分别建立多模式ENSO预测结果智能会商系统,对多模式预测结果进行集合订正,并给出各模式预测结果在智能会商系统中的特征重要性。该智能会商系统模拟了ENSO预测会商过程,实现了读取各模式预测结果、训练模型、给出预测结论及预测依据、预测结果可视化等流程的自动化,同时实现了智能调参的功能。Bansed on the concept of multi-model ensemble forecasting,this study establishes an intelligent consultation system for multi-model intelligent consultation system of ENSO prediction using the interpretable machine learning method named decision tree algorithm.The hyper parameters of four decision tree models of GBDT based on Boosting,XGBoost,lightGBM and Random Forest(RF)based on Bagging are optimized and adjusted by using two hyper parameter adjustment methods of random search cross-validation and grid search cross-validation.The intelligent consultation system of multi-model ENSO prediction results is established according to different prediction leading time,which makes integrated correction on the multi-model prediction results and provides the feature importance of the prediction result of each model in the intelligent consultation system.The intelligent consultation system simulates the consultation process of ENSO prediction,which realizes the automation of the processes of reading the prediction results of each model,training the model,giving the prediction conclusion and prediction basis and the visualization of the prediction results,and realizes the function of intelligent parameter tuning.The intelligent consultation system collectively revises the multi-modal ENSO prediction results.The results show that machine learning also has some advantages in the multi-modal result consultation,which provide a reference for consultations of ENSO prediction in the future.

关 键 词：ENSO 可解释机器学习 多模式 智能会商 

分 类 号：P456[天文地球—大气科学及气象学]