Scholarlink

<p>Understanding long-term wind variability in monsoon-dominated coastal regions is essential for climate risk assessment, environmental management, and renewable energy planning. However, most previous studies have focused primarily on short-term forecasting and have not adequately examined multi-decadal wind dynamics under non-stationary climatic conditions. This study presents a comprehensive spatiotemporal analysis of coastal and near-coastal wind behavior in southern Bangladesh using a multi-model machine-learning framework applied to long-term observational wind records (1971–2025) from 16 meteorological stations. A rigorous chronological modeling framework was implemented using separate training (1971–2010), validation (2011–2020), and independent test (2021–2025) periods to ensure realistic evaluation under evolving climate conditions. Twelve physically meaningful predictors were engineered, including temporal trends, harmonic seasonal indicators, station-specific climatological features, and lag-based wind predictors. Six machine-learning models—Linear Regression, Ridge Regression, Decision Tree, Random Forest, Extra Trees, and Gradient Boosting—were evaluated using multiple statistical performance metrics. Results reveal a statistically significant decline in wind speed at most stations, with reductions of up to 55%, indicating pronounced wind stilling in recent decades. Model performance was strongly influenced by the degree of climatic non-stationarity: stations with relatively stable wind regimes maintained high predictive skill, whereas stations experiencing severe wind stilling showed substantial performance degradation, including negative test-period R<sup>2</sup> values. Ensemble-based approaches generally outperformed linear models under relatively stable conditions but struggled to generalize under strong regime shifts. Nevertheless, persistent seasonal cycles enabled partial retention of predictive skill despite long-term declines in wind trends. These findings demonstrate that predictive performance in long-term environmental modeling is fundamentally constrained by climate non-stationarity rather than algorithmic complexity alone. The study offers new insights into how seasonal variability interacts with long-term wind trends in monsoon-dominated coastal systems. It presents a robust methodological framework for climate-aware machine learning applications that support coastal planning, environmental management, and hazard assessment.</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

Seasonality and wind stilling in a changing coastal climate: machine learning insights from a multi-decadal analysis in Bangladesh

  • 期刊论文

  • 发表日期：

    2026年7月2日

查看全文
  • Hasan Ahamed Alif,
  • Md. Jisan Mashrafi,
  • Avijit Deb Nath,
  • Suhas Singha,
  • Jeba Maliha,
  • Md. Sazidul Islam,
  • Mezabahnur Masum,
  • Md. Zia Uddin Raihan,
  • Muttaki Islam Bismoy

摘要

Understanding long-term wind variability in monsoon-dominated coastal regions is essential for climate risk assessment, environmental management, and renewable energy planning. However, most previous studies have focused primarily on short-term forecasting and have not adequately examined multi-decadal wind dynamics under non-stationary climatic conditions. This study presents a comprehensive spatiotemporal analysis of coastal and near-coastal wind behavior in southern Bangladesh using a multi-model machine-learning framework applied to long-term observational wind records (1971–2025) from 16 meteorological stations. A rigorous chronological modeling framework was implemented using separate training (1971–2010), validation (2011–2020), and independent test (2021–2025) periods to ensure realistic evaluation under evolving climate conditions. Twelve physically meaningful predictors were engineered, including temporal trends, harmonic seasonal indicators, station-specific climatological features, and lag-based wind predictors. Six machine-learning models—Linear Regression, Ridge Regression, Decision Tree, Random Forest, Extra Trees, and Gradient Boosting—were evaluated using multiple statistical performance metrics. Results reveal a statistically significant decline in wind speed at most stations, with reductions of up to 55%, indicating pronounced wind stilling in recent decades. Model performance was strongly influenced by the degree of climatic non-stationarity: stations with relatively stable wind regimes maintained high predictive skill, whereas stations experiencing severe wind stilling showed substantial performance degradation, including negative test-period R2 values. Ensemble-based approaches generally outperformed linear models under relatively stable conditions but struggled to generalize under strong regime shifts. Nevertheless, persistent seasonal cycles enabled partial retention of predictive skill despite long-term declines in wind trends. These findings demonstrate that predictive performance in long-term environmental modeling is fundamentally constrained by climate non-stationarity rather than algorithmic complexity alone. The study offers new insights into how seasonal variability interacts with long-term wind trends in monsoon-dominated coastal systems. It presents a robust methodological framework for climate-aware machine learning applications that support coastal planning, environmental management, and hazard assessment.