Scholarlink

<p>This study proposes a hybrid deep learning model combining a one-dimensional Convolutional Neural Network (CNN-1D) and a Bidirectional Long Short-Term Memory (BiLSTM) network for short-term wind power prediction. Hourly averages of atmospheric, environmental, and power variables collected from wind farms in Germany are utilised as input. Model performance is evaluated using Root Mean Square Error (RMSE) and Mean Absolute Error (MAE) metrics. To enhance prediction accuracy, the raw data undergoes pre-processing. Two different training/testing data split scenarios–70%-30% and 50%-50%–are investigated. Predictions ranging from one to twelve hours ahead are analysed in detail, and the proposed model is compared with other approaches. Furthermore, the results are benchmarked against previous studies that employed the same dataset. While the compared models exhibit an increase in forecast error of up to 54 times as the prediction horizon extends, the proposed model shows only a 1.54-fold difference between the maximum and minimum errors from the first to the 12th hour. Overall, the proposed hybrid system demonstrates superior predictive capability compared to the methods it was compared against.</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

Improved performance of short-term wind power forecasting system using a hybrid deep learning model with data pre-processing

  • 期刊论文

  • 发表日期：

    2026年7月24日

查看全文
  • Huseyin Oktay Erkol,
  • Abdulsamed Tabak

摘要

This study proposes a hybrid deep learning model combining a one-dimensional Convolutional Neural Network (CNN-1D) and a Bidirectional Long Short-Term Memory (BiLSTM) network for short-term wind power prediction. Hourly averages of atmospheric, environmental, and power variables collected from wind farms in Germany are utilised as input. Model performance is evaluated using Root Mean Square Error (RMSE) and Mean Absolute Error (MAE) metrics. To enhance prediction accuracy, the raw data undergoes pre-processing. Two different training/testing data split scenarios–70%-30% and 50%-50%–are investigated. Predictions ranging from one to twelve hours ahead are analysed in detail, and the proposed model is compared with other approaches. Furthermore, the results are benchmarked against previous studies that employed the same dataset. While the compared models exhibit an increase in forecast error of up to 54 times as the prediction horizon extends, the proposed model shows only a 1.54-fold difference between the maximum and minimum errors from the first to the 12th hour. Overall, the proposed hybrid system demonstrates superior predictive capability compared to the methods it was compared against.