Scholarlink

Reliable probabilistic wind forecasts are essential for managing renewable energy systems. This study compares Bayesian Additive Regression Trees (BART), Gaussian Process Regression (GPR), and Quantile Random Forests (QRF) under at-site versus regional (pooled) modeling regimes, using 21 years of daily wind data from eleven NSW and QLD stations. Models are evaluated for point accuracy (RMSE), probabilistic skill (CRPS), and interval coverage. Regional pooling marginally affects QRF’s RMSE (+0.006) but increases RMSE for BART (+0.063) and GPR (+0.18). GPR maintains superior coverage (0.959 at-site, 0.945 regional) and lowest post-pooling CRPS (1.40), while QRF’s intervals remain under-dispersed (~0.70 coverage). BART’s coverage decreases sharply under pooling (0.95 to 0.69). Variable importance consistently ranks pressure and temperature highest. We provide clear guidelines for operational deployment, highlighting GPR’s strengths for calibrated probabilistic forecasting in pooled environments.

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

Regional vs. At-Site Probabilistic Wind Forecasting in NSW and QLD: A Comparative Study of Bayesian Adaptive Regression Trees, Gaussian Process Regression, and Quantile Random Forests

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Khaled Haddad,
  • Ataur Rahman

摘要

Reliable probabilistic wind forecasts are essential for managing renewable energy systems. This study compares Bayesian Additive Regression Trees (BART), Gaussian Process Regression (GPR), and Quantile Random Forests (QRF) under at-site versus regional (pooled) modeling regimes, using 21 years of daily wind data from eleven NSW and QLD stations. Models are evaluated for point accuracy (RMSE), probabilistic skill (CRPS), and interval coverage. Regional pooling marginally affects QRF’s RMSE (+0.006) but increases RMSE for BART (+0.063) and GPR (+0.18). GPR maintains superior coverage (0.959 at-site, 0.945 regional) and lowest post-pooling CRPS (1.40), while QRF’s intervals remain under-dispersed (~0.70 coverage). BART’s coverage decreases sharply under pooling (0.95 to 0.69). Variable importance consistently ranks pressure and temperature highest. We provide clear guidelines for operational deployment, highlighting GPR’s strengths for calibrated probabilistic forecasting in pooled environments.