Scholarlink

Purpose <p>Precise precipitation forecasting is necessary for both the management of water resources and preparedness for severe weather events (especially in climate sensitive regions such as the Mediterranean). The aim of this research project is to design and implement an auditable, two-stage machine-learning based approach using transparent CatBoost regression models to generate quantitative precipitation forecasts (QPF), and LightGBM classification models to predict extreme weather event alarms.</p> Methods <p>Utilizing a complete 24-year archive of daily weather data (2001–2024) from Köyceğiz, Turkey, this research provides a dual-task workflow. The first task is an optimized CatBoost regression model that provides quantitative precipitation forecasts (QPF). The second task is a LightGBM classification model that identifies extreme anomaly events (above the 95th percentile), utilizing F1-score as an optimization method to account for class imbalance.</p> Results <p>Beginning with an evaluation using a long-term and wide-ranging dataset of 24&#xa0;years of daily values (2001–2024) CatBoost was able to explain 83% of overall rainfall variability and provided generalized information on rainfall regimes. During the same period, the LightGBM classifier successfully identified 75% of extreme precipitation events (defined by the 95th percentile threshold of 15.24&#xa0;mm/day), and its area under the ROC curve exceeded 0.97; however, the very high separability of the two classes is due at least in part to the substantial amount of temporal autocorrelation present in this region’s hydroclimatological data. In addition, examination of the SHAP values demonstrated that, regardless of whether predicting the total amount of precipitation or the probability of an extreme event, the top driving variables were land surface characteristics (soil moisture) and relative humidity.</p> Conclusions <p>The use of multiple variables, in combination with lag elements, significantly improves predictive capabilities over that of a baseline single variable. While this research was conducted as a single location, the methodology and the complete listing of attributes are adaptable to reproduce and subsequently test at other locations with similar Mediterranean climates.</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

High-accuracy day-ahead precipitation forecasting with an interpretable CatBoost regressor and a LightGBM-based extreme-event alerting pipeline

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月1日

查看全文
  • Tevfik Denizhan Muftuoglu

摘要

Purpose

Precise precipitation forecasting is necessary for both the management of water resources and preparedness for severe weather events (especially in climate sensitive regions such as the Mediterranean). The aim of this research project is to design and implement an auditable, two-stage machine-learning based approach using transparent CatBoost regression models to generate quantitative precipitation forecasts (QPF), and LightGBM classification models to predict extreme weather event alarms.

Methods

Utilizing a complete 24-year archive of daily weather data (2001–2024) from Köyceğiz, Turkey, this research provides a dual-task workflow. The first task is an optimized CatBoost regression model that provides quantitative precipitation forecasts (QPF). The second task is a LightGBM classification model that identifies extreme anomaly events (above the 95th percentile), utilizing F1-score as an optimization method to account for class imbalance.

Results

Beginning with an evaluation using a long-term and wide-ranging dataset of 24 years of daily values (2001–2024) CatBoost was able to explain 83% of overall rainfall variability and provided generalized information on rainfall regimes. During the same period, the LightGBM classifier successfully identified 75% of extreme precipitation events (defined by the 95th percentile threshold of 15.24 mm/day), and its area under the ROC curve exceeded 0.97; however, the very high separability of the two classes is due at least in part to the substantial amount of temporal autocorrelation present in this region’s hydroclimatological data. In addition, examination of the SHAP values demonstrated that, regardless of whether predicting the total amount of precipitation or the probability of an extreme event, the top driving variables were land surface characteristics (soil moisture) and relative humidity.

Conclusions

The use of multiple variables, in combination with lag elements, significantly improves predictive capabilities over that of a baseline single variable. While this research was conducted as a single location, the methodology and the complete listing of attributes are adaptable to reproduce and subsequently test at other locations with similar Mediterranean climates.