Scholarlink

This study aims to comprehensively explore and refine the evaluation of predictive model reliability for firefighting interventions. We have acknowledged that while predictive models may aid in anticipating interventions, they may lack reliability in critical scenarios where prioritization is essential. Thus, we seek to develop a tool that provides a confidence index for predicted outcomes. Given the inherent urgency in firefighting situations, our inquiry examines how predictive model reliability evaluation influences firefighting interventions over nine years, from 2015 to 2024. The annual record of firefighter interventions in France reveals particularly long response times in certain departments, underscoring the critical nature of our research. Our methodology entails data preparation, thorough contextual analysis, and the application of four machine learning approaches. We combined XGBoost and LightGBM models to measure their proximity, utilized Quantile regression to gauge the confidence interval width [0.05, 0.95], applied the NgBoost model to measure the same confidence interval from the provided distribution, and finally employed Calibrated probability through XGBoost classification to measure the most probable class. We assessed reliability through multiple perspectives, including overall reliability, seasons, and event peaks. The findings of our study unveiled complementarity among these approaches in the aforementioned perspectives. Specifically, we found that the Quantile regression approach offers stable overall performance and a high average reliability score. Additionally, this approach and the XGBoost versus LightGBM model approach are advantageous for the winter season, where reliability is more stable. Conversely, the NgBoost model is more suitable for the summer season. Regarding event peaks, the results demonstrated that the contribution of NgBoost and Calibrated probability was a prudent choice to minimize damages. In conclusion, this study provides a practical approach to anticipate and optimize resource management, improve firefighter response times, and contribute to saving lives by mitigating intervention failures during major incidents.

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

Reliability of Firefighter Prediction Models Across Time and Peaks

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Naoufal Sirri,
  • Christophe Guyeux

摘要

This study aims to comprehensively explore and refine the evaluation of predictive model reliability for firefighting interventions. We have acknowledged that while predictive models may aid in anticipating interventions, they may lack reliability in critical scenarios where prioritization is essential. Thus, we seek to develop a tool that provides a confidence index for predicted outcomes. Given the inherent urgency in firefighting situations, our inquiry examines how predictive model reliability evaluation influences firefighting interventions over nine years, from 2015 to 2024. The annual record of firefighter interventions in France reveals particularly long response times in certain departments, underscoring the critical nature of our research. Our methodology entails data preparation, thorough contextual analysis, and the application of four machine learning approaches. We combined XGBoost and LightGBM models to measure their proximity, utilized Quantile regression to gauge the confidence interval width [0.05, 0.95], applied the NgBoost model to measure the same confidence interval from the provided distribution, and finally employed Calibrated probability through XGBoost classification to measure the most probable class. We assessed reliability through multiple perspectives, including overall reliability, seasons, and event peaks. The findings of our study unveiled complementarity among these approaches in the aforementioned perspectives. Specifically, we found that the Quantile regression approach offers stable overall performance and a high average reliability score. Additionally, this approach and the XGBoost versus LightGBM model approach are advantageous for the winter season, where reliability is more stable. Conversely, the NgBoost model is more suitable for the summer season. Regarding event peaks, the results demonstrated that the contribution of NgBoost and Calibrated probability was a prudent choice to minimize damages. In conclusion, this study provides a practical approach to anticipate and optimize resource management, improve firefighter response times, and contribute to saving lives by mitigating intervention failures during major incidents.