Scholarlink

<p>Monthly precipitation forecasting in Iran and its neighboring regions, due to its complex and diverse climate, always faces many challenges. This article pursues three objectives. The first is to improve the monthly precipitation forecasting of Iran and its surrounding areas using ensemble machine learning (ML) methods. The second objective is to use SHAP (SHapley Additive exPlanations) in order to explain and interpret the outputs of used ML algorithms. Finally, the third objective is to rank the climate models based on the MCDM algorithm Characteristic Objects Method (COMET). For this purpose, the outputs of four different climate models are considered for post-processing and also as inputs to eight machine learning algorithms. These climate models include ECMWF, and three models from the NMME phase2 project including, COLA-RSMAS-CCSM4 (CCSM4), NASA-GEOSS2S (NASA), and GFDL-SPEAR (SPEAR). Among the machine learning methods, in addition to Artificial Neural Network (ANN), k-nearest neighbors (KNN), Random Forest (RF), Histogram-based Gradient Boosting (Hist), XGBoost (XBG) and Gradient Boosting (GB), the final ensemble learning method Stacking (Stack) was also used. ERA5 dataset was used to train the eight ML algorithms during the time period 1993-2024. Results indicated that for most of the 12 months, Stack and ANN algorithms outperformed the other ones. Also, among the four climate models, ECMWF and SPEAR outperformed the others. Finally, a ranking of the models is proposed using the MCDM algorithm Characteristic Objects Method (COMET), which indicates the best model for each month of the year. Based on the results, among the climate models, ECMWF had the best performance for all months except for June and November, in which the Spear model had the best performance. Also, among the ML algorithms, Stacking had the best performance for all months except for February, July, September and December, in which ANN algorithm had the best performance.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Multi-model ensemble forecasting of precipitation: explainable machine learning and ranking of climate models based on characteristic objects method

  • 期刊论文

  • 发表日期：

    2026年1月17日

查看全文
  • Abouzar Mehraban,
  • Morteza Pakdaman,
  • Shaoqi Gong,
  • Majid Habibi Nokhandan

摘要

Monthly precipitation forecasting in Iran and its neighboring regions, due to its complex and diverse climate, always faces many challenges. This article pursues three objectives. The first is to improve the monthly precipitation forecasting of Iran and its surrounding areas using ensemble machine learning (ML) methods. The second objective is to use SHAP (SHapley Additive exPlanations) in order to explain and interpret the outputs of used ML algorithms. Finally, the third objective is to rank the climate models based on the MCDM algorithm Characteristic Objects Method (COMET). For this purpose, the outputs of four different climate models are considered for post-processing and also as inputs to eight machine learning algorithms. These climate models include ECMWF, and three models from the NMME phase2 project including, COLA-RSMAS-CCSM4 (CCSM4), NASA-GEOSS2S (NASA), and GFDL-SPEAR (SPEAR). Among the machine learning methods, in addition to Artificial Neural Network (ANN), k-nearest neighbors (KNN), Random Forest (RF), Histogram-based Gradient Boosting (Hist), XGBoost (XBG) and Gradient Boosting (GB), the final ensemble learning method Stacking (Stack) was also used. ERA5 dataset was used to train the eight ML algorithms during the time period 1993-2024. Results indicated that for most of the 12 months, Stack and ANN algorithms outperformed the other ones. Also, among the four climate models, ECMWF and SPEAR outperformed the others. Finally, a ranking of the models is proposed using the MCDM algorithm Characteristic Objects Method (COMET), which indicates the best model for each month of the year. Based on the results, among the climate models, ECMWF had the best performance for all months except for June and November, in which the Spear model had the best performance. Also, among the ML algorithms, Stacking had the best performance for all months except for February, July, September and December, in which ANN algorithm had the best performance.