An Intelligent Consulation System About Multi-Model Forecasts Of ENSO Based On Interpretable Machine Learning

Forecasts of ENSO mainly rely on dynamic models and statistical models.The continuous improvement of computer and machine learning algorithms provide a new tool for E N S O prediction.Research results at home and abroad show that the accuracy of weather forecasts,climate prediction,and model result correction can achieve or surpass traditional methods,with the use of machine learning methods.Based on the idea of forecast consultation,this research uses one kind of interpretable machine learning methods--decision tree algorithms to establish an intelligent consultation system for multi-model forecasts of ENSO.In this study,four decision tree model methods are used,including GBDT,XGBoost,light GBM and RF,which based on Boosting or Bagging.Combined with two hyperparameter adjustment methods,random search cross-validation and grid search cross-validation,this study optimizes and adjusts the hyperparameters of tree models to establish an intelligent consultation system about multi-mode E N S O predicts.This system makes an integrated revised experiment about the multi-mode prediction results,and provides feature importance of every model in the intelligent consultation system.Finally,the results of decision tree methods are compared with the results of the traditional methods.The main results are as follows:In terms of feature importance,different decision tree models show different preferences.Overall,when lead time is shorter,the dynamic model is more important;when lead time is longer,the statistical model is more important.This is the same as forecasting skills of dynamic models and statistical models at different lead times.In terms of time series diagrams,the prediction results of the four tree model methods all have the following characteristics: at seasonal month 1 ? 3 lead time the phase and intensity are basically the same as the label values;at 4 ? 6 lead time the deviation between forecasts and label values is small,and there is an over-fitting problem;at 7 ? 9 lead time deviation are relatively large.Generally speaking,with the increase of lead time,the degree of phase lag gradually increases,and the deviation of intensity becomes larger and larger.In terms of evaluation indicators of forecasting skill,different decision tree models show different forecasting skill at different lead times.At seasonal month 1 ?3 lead time,forecasting skill of integrated revised model based on GBDT are the best;at 4 ? 5,XGBoost are the best;at 6 ? 7,light GBM are the best;at 8 ? 9,RF are the best.In terms of the ensemble average results,the evaluation indicators of forecasting skill of the decision tree models at different lead times are all equal to or better than the dynamic models and the statistical models.The innovation of this research are reflected in the using of interpretable machine learning methods for multi-model ensemble correction,and an attempt to explain the modeling and prediction process of machine learning algorithms.The feature importance of the prediction results of each model in this system at different lead times are given.This job may provide a reference for future E N S O forecast consultation.