Scholarlink

<p>Lightning strikes are natural phenomena with significant implications for human safety, infrastructure, and the environment. Predicting lightning events is crucial for mitigating risks, improving public safety, and optimizing the management of electrical grids and aviation operations. Study area has recently witnessed a sharp increase in lightning hazards during the rainy season, with over 20,000 lightning-related incidents recorded in 2023 alone, leading to significant loss of life and property in multiple districts. However, predicting x lightning’s exact location, timing, and intensity remains challenging due to the complex interplay of thunderstorms and atmospheric dynamics. We proposed an integrated framework combining geospatial techniques with an ensemble-based architecture for lightning flash prediction to address these challenges. As the first-level individual learners, the ensemble strategy combines Random Forests (RF), Logistic Regression (LR), and XGBoost. To increase precision, their predictions are integrated using a voting classifier at the second level. The model leverages the geospatial and atmospheric coordinates to accurately forecast lightning strikes, including longitudinal and latitudinal coordinates, radiance, milliseconds, groups, events, etc. The test results reveal that the ensemble model achieves good performance, getting an area under the curve (AUC) score of 88%, which is higher than the performance of individual learners such as RF (83%), LR (79%), and XGBoost (84%). The proposed model also obtains 0.87, 0.86, and 0.86 for precision, recall, and F1-scores, respectively, all of which are greater than those of individual learners. Conversely, the research highlights that geospatial data enhances the targeted prediction and enables the definition of specific areas and time intervals with the highest risk expectancy. From this perspective, we can see geospatial data being utilized in conjunction with retrospective artificial intelligence approaches for lightning forecasting, which provides clear advantages regarding risk and safety management practice. The results provide a new data-driven approach in the management and planning of measures aiming at lightning threats and call rather for applying more complex ensemble-based models in key meteorological parameters.</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

Enhancing lightning strike prediction: leveraging geospatial data and ensemble machine learning for risk management

  • 期刊论文

  • 发表日期：

    2025年7月29日

查看全文
  • Partha Sarathi Mishra,
  • Debabrata Nandi,
  • Rakesh Ranjan Thakur,
  • Sujit Kumar Roy,
  • Roshan Beuria,
  • Subhasmita Das,
  • Hela Elmannai,
  • Aqil Tariq

摘要

Lightning strikes are natural phenomena with significant implications for human safety, infrastructure, and the environment. Predicting lightning events is crucial for mitigating risks, improving public safety, and optimizing the management of electrical grids and aviation operations. Study area has recently witnessed a sharp increase in lightning hazards during the rainy season, with over 20,000 lightning-related incidents recorded in 2023 alone, leading to significant loss of life and property in multiple districts. However, predicting x lightning’s exact location, timing, and intensity remains challenging due to the complex interplay of thunderstorms and atmospheric dynamics. We proposed an integrated framework combining geospatial techniques with an ensemble-based architecture for lightning flash prediction to address these challenges. As the first-level individual learners, the ensemble strategy combines Random Forests (RF), Logistic Regression (LR), and XGBoost. To increase precision, their predictions are integrated using a voting classifier at the second level. The model leverages the geospatial and atmospheric coordinates to accurately forecast lightning strikes, including longitudinal and latitudinal coordinates, radiance, milliseconds, groups, events, etc. The test results reveal that the ensemble model achieves good performance, getting an area under the curve (AUC) score of 88%, which is higher than the performance of individual learners such as RF (83%), LR (79%), and XGBoost (84%). The proposed model also obtains 0.87, 0.86, and 0.86 for precision, recall, and F1-scores, respectively, all of which are greater than those of individual learners. Conversely, the research highlights that geospatial data enhances the targeted prediction and enables the definition of specific areas and time intervals with the highest risk expectancy. From this perspective, we can see geospatial data being utilized in conjunction with retrospective artificial intelligence approaches for lightning forecasting, which provides clear advantages regarding risk and safety management practice. The results provide a new data-driven approach in the management and planning of measures aiming at lightning threats and call rather for applying more complex ensemble-based models in key meteorological parameters.