Scholarlink

This study develops a comprehensive framework for quantifying prediction intervals (PIs) in hour-ahead wind energy forecasting by combining Long Short-Term Memory (LSTM) networks and Bayesian optimization. The inherent uncertainty in wind energy output necessitates reliable forecasting techniques, and the Lower Upper Bound Estimation (LUBE) method is employed to generate PIs that effectively capture the variability in predictions. A dynamic composite loss function is proposed, balancing Prediction Interval Coverage Probability (PICP) and Prediction Interval Width (PIW), to ensure a robust trade-off between interval reliability and precision. To enhance the model’s performance, Bayesian optimization is applied to fine-tune critical hyperparameters, including the lambda parameter, which controls the trade-off between PIW and PICP. Results reveal the model’s capability to achieve a moderate PIW of 98.59 kWh. The midpoint predictions demonstrate an RMSE of 92.77 kWh, validating the model’s predictive accuracy. These findings underline the potential of the proposed approach to improve interval optimization and enhance reliability in wind energy forecasting. The study lays a foundation for further advancements in wind energy forecasting, emphasizing the importance of balancing precision and reliability in interval predictions.

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

LSTM and Bayesian Computation in Uncertainty Quantification for Wind Energy Forecasting

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Mohammed Al-Mahmodi,
  • Hanan Mansi,
  • Changqing Cheng,
  • Frederic Heliodore,
  • Yong Wang

摘要

This study develops a comprehensive framework for quantifying prediction intervals (PIs) in hour-ahead wind energy forecasting by combining Long Short-Term Memory (LSTM) networks and Bayesian optimization. The inherent uncertainty in wind energy output necessitates reliable forecasting techniques, and the Lower Upper Bound Estimation (LUBE) method is employed to generate PIs that effectively capture the variability in predictions. A dynamic composite loss function is proposed, balancing Prediction Interval Coverage Probability (PICP) and Prediction Interval Width (PIW), to ensure a robust trade-off between interval reliability and precision. To enhance the model’s performance, Bayesian optimization is applied to fine-tune critical hyperparameters, including the lambda parameter, which controls the trade-off between PIW and PICP. Results reveal the model’s capability to achieve a moderate PIW of 98.59 kWh. The midpoint predictions demonstrate an RMSE of 92.77 kWh, validating the model’s predictive accuracy. These findings underline the potential of the proposed approach to improve interval optimization and enhance reliability in wind energy forecasting. The study lays a foundation for further advancements in wind energy forecasting, emphasizing the importance of balancing precision and reliability in interval predictions.