<p>Reliable monthly rainfall prediction remains a critical challenge in climate science, with significant implications for agriculture, water resource management, and disaster mitigation. This study introduces an Enhanced Physical Factor Set (EPFS)-based Ensemble Machine Learning Model (EPFM) to improve June rainfall forecasting in Hunan Province, China. The model integrates Niño-based climate signals and large-scale circulation patterns, to better capture the spatial and temporal variability in precipitation. To address the challenge of limited training data, a systematic data expansion strategy is proposed, increasing the sample size 125-times through controlling boundaries of empirical orthogonal function (EOF) analysis for selecting the factors. An ensemble Support Vector Regression (SVR) model is developed by cross-validation techniques, leveraging Probability Density Correction (PDC) to refine predictions. The model is trained using monthly geopotential heights (500 hPa), velocity potential (850 hPa and 200 hPa) from NCEP/NCAR reanalysis data, and rainfall observations from 97 meteorological stations over Hunan. Additionally, 10-day lead rainfall forecasts from ECMWF’s S2S model are used for performance evaluation. The results demonstrate that EPFM significantly outperforms traditional numerical weather models, achieving an average prediction skill (<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\({\text{P}}_{\text{s}}\)</EquationSource> </InlineEquation>) score of 76 in independent validation (2016–2022), compared to 67 for numerical models. By integrating physical predictors with machine learning techniques, this study establishes a framework for improving regional monthly rainfall predictions. These findings highlight EPFM as a robust tool for disaster preparedness, water resource management, and climate resilience, with broader applicability for enhancing early warning systems in regions prone to extreme weather events.</p> Graphical Abstract <p>Based on the graphical framework, this study was conducted to enhance monthly rainfall forecasting in Hunan Province by integrating physical climate dynamics with ensemble machine learning (ML). The work captures the complex relationships between large-scale atmospheric predictors—including geopotential heights, velocity potential, and Niño-based signals—and regional precipitation variability. The graph shows the technical path-solving process of this method. To address data scarcity, a systematic sample expansion method was applied, increasing training samples 125-times through empirical orthogonal function (EOF)-guided factor selection. Then, an ensemble prediction model was developed using Support Vector Regression (SVR), optimized via cross-validation and refined with Probability Density Correction (PDC) to improve forecast reliability. The upper part graph represents the training process of the SVR model, where <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\({\text{X}}_{0}\)</EquationSource> </InlineEquation> denotes the predictors, and Y represents the training set of targets. <InlineEquation ID="IEq3"> <EquationSource Format="TEX">\({\text{F}}^{\text{C}}\left(\cdot \:\right)\)</EquationSource> </InlineEquation> is the output of Probability Density Correction (PDC) model, and <InlineEquation ID="IEq4"> <EquationSource Format="TEX">\({\text{Y}}_{\text{P}\text{D}\text{C}}\)</EquationSource> </InlineEquation> represents the the training set of targets after PDC. The model was trained on NCEP/NCAR reanalysis data and station observations (1981–2015). For June rainfall forecasting, the factors are usually obtained a month lead with factors in April to input the model. Then the model was validated against 10-days lead ECMWF S2S forecasts (2016–2022). The lower part graph is the prediction period (also known as independent test period). The 2016–2022 NCEP/NCAR Reanalysis Data <InlineEquation ID="IEq5"> <EquationSource Format="TEX">\({\text{X}}_{1}\)</EquationSource> </InlineEquation> is regarded as the input used to obtain the output <InlineEquation ID="IEq6"> <EquationSource Format="TEX">\(\text{F}\left({\text{X}}_{1}\right)\)</EquationSource> </InlineEquation> as a result derived by trained cross-validation models. Thus <InlineEquation ID="IEq7"> <EquationSource Format="TEX">\({\text{Y}}_{\text{p}\text{r}\text{e}}\)</EquationSource> </InlineEquation> represents results of the independent test. Results demonstrate that the proposed Enhanced Physical Factor-based Model (EPFM) achieves superior predictive skill (score: 76) compared to numerical weather models (score: 67), highlighting its effectiveness for operational rainfall forecasting. Key findings reveal the critical role of integrating physical predictors with ML techniques to capture spatiotemporal rainfall patterns. This framework offers actionable insights for disaster preparedness, water resource management, and climate resilience in regions vulnerable to extreme weather. The study underscores EPFM as a innovative tool for advancing early warning systems and mitigating hydroclimatic risks.</p>

Integrating Physical Dynamics into Ensemble ML for Improved Monthly Rainfall Forecasting

期刊论文
发表日期：

2025年6月20日

查看全文

Yuxing Yang,
Guirong Tan,
Ziqi Shen,
Yi Zhang,
Qiming Fei,
Xiaoyun Liu,
Muhammad Ahmad Dogar

摘要

Reliable monthly rainfall prediction remains a critical challenge in climate science, with significant implications for agriculture, water resource management, and disaster mitigation. This study introduces an Enhanced Physical Factor Set (EPFS)-based Ensemble Machine Learning Model (EPFM) to improve June rainfall forecasting in Hunan Province, China. The model integrates Niño-based climate signals and large-scale circulation patterns, to better capture the spatial and temporal variability in precipitation. To address the challenge of limited training data, a systematic data expansion strategy is proposed, increasing the sample size 125-times through controlling boundaries of empirical orthogonal function (EOF) analysis for selecting the factors. An ensemble Support Vector Regression (SVR) model is developed by cross-validation techniques, leveraging Probability Density Correction (PDC) to refine predictions. The model is trained using monthly geopotential heights (500 hPa), velocity potential (850 hPa and 200 hPa) from NCEP/NCAR reanalysis data, and rainfall observations from 97 meteorological stations over Hunan. Additionally, 10-day lead rainfall forecasts from ECMWF’s S2S model are used for performance evaluation. The results demonstrate that EPFM significantly outperforms traditional numerical weather models, achieving an average prediction skill ( \({\text{P}}_{\text{s}}\) ) score of 76 in independent validation (2016–2022), compared to 67 for numerical models. By integrating physical predictors with machine learning techniques, this study establishes a framework for improving regional monthly rainfall predictions. These findings highlight EPFM as a robust tool for disaster preparedness, water resource management, and climate resilience, with broader applicability for enhancing early warning systems in regions prone to extreme weather events.

Graphical Abstract

Based on the graphical framework, this study was conducted to enhance monthly rainfall forecasting in Hunan Province by integrating physical climate dynamics with ensemble machine learning (ML). The work captures the complex relationships between large-scale atmospheric predictors—including geopotential heights, velocity potential, and Niño-based signals—and regional precipitation variability. The graph shows the technical path-solving process of this method. To address data scarcity, a systematic sample expansion method was applied, increasing training samples 125-times through empirical orthogonal function (EOF)-guided factor selection. Then, an ensemble prediction model was developed using Support Vector Regression (SVR), optimized via cross-validation and refined with Probability Density Correction (PDC) to improve forecast reliability. The upper part graph represents the training process of the SVR model, where \({\text{X}}_{0}\) denotes the predictors, and Y represents the training set of targets. \({\text{F}}^{\text{C}}\left(\cdot \:\right)\) is the output of Probability Density Correction (PDC) model, and \({\text{Y}}_{\text{P}\text{D}\text{C}}\) represents the the training set of targets after PDC. The model was trained on NCEP/NCAR reanalysis data and station observations (1981–2015). For June rainfall forecasting, the factors are usually obtained a month lead with factors in April to input the model. Then the model was validated against 10-days lead ECMWF S2S forecasts (2016–2022). The lower part graph is the prediction period (also known as independent test period). The 2016–2022 NCEP/NCAR Reanalysis Data \({\text{X}}_{1}\) is regarded as the input used to obtain the output \(\text{F}\left({\text{X}}_{1}\right)\) as a result derived by trained cross-validation models. Thus \({\text{Y}}_{\text{p}\text{r}\text{e}}\) represents results of the independent test. Results demonstrate that the proposed Enhanced Physical Factor-based Model (EPFM) achieves superior predictive skill (score: 76) compared to numerical weather models (score: 67), highlighting its effectiveness for operational rainfall forecasting. Key findings reveal the critical role of integrating physical predictors with ML techniques to capture spatiotemporal rainfall patterns. This framework offers actionable insights for disaster preparedness, water resource management, and climate resilience in regions vulnerable to extreme weather. The study underscores EPFM as a innovative tool for advancing early warning systems and mitigating hydroclimatic risks.