Scholarlink

<p>Accurate load prediction in combined-cycle power plants is critical for optimizing unit operations, reducing energy waste, and improving grid stability. However, traditional forecasting methods often fail to address the dynamic nature of gas-steam cycles and the time lag inherent in load dispatching. To bridge this gap, this study proposes a novel hybrid neural network model that integrates long short-term memory (LSTM) with an attention mechanism (AM) to enhance prediction accuracy and operational efficiency. Using historical operating data from three gas-steam combined cycle units, we identify key influencing parameters (mass airflow, atmospheric temperature, compressor inlet temperature, and steam outlet flow) through correlation analysis. The LSTM-AM model leverages temporal dependencies in time-series data while adaptively weighting influential inputs via attentive learning. Trained with the Adam optimizer to minimize mean squared error, the model achieves superior performance, with a mean absolute percentage error (MAPE) of 0.927 % and a root mean squared error (RMSE) of 4.030 MW, outperforming baseline RNN, GRU, and standalone LSTM models. Furthermore, the model enables proactive load scheduling by predicting hourly demand, reducing unnecessary unit startups/shutdowns, and cutting energy waste. This work contributes a practical framework for power plant operators to optimize dispatch strategies, demonstrating the potential of hybrid deep learning in advancing energy efficiency and operational decision-making.</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

Optimizing combined-cycle power plant operations using an LSTM-attention hybrid model for load forecasting

  • 期刊论文

  • 发表日期：

    2025年9月19日

查看全文
  • Anping Wan,
  • Chenyu Du,
  • Khalil AL-Bukhaiti,
  • Peng Chen

摘要

Accurate load prediction in combined-cycle power plants is critical for optimizing unit operations, reducing energy waste, and improving grid stability. However, traditional forecasting methods often fail to address the dynamic nature of gas-steam cycles and the time lag inherent in load dispatching. To bridge this gap, this study proposes a novel hybrid neural network model that integrates long short-term memory (LSTM) with an attention mechanism (AM) to enhance prediction accuracy and operational efficiency. Using historical operating data from three gas-steam combined cycle units, we identify key influencing parameters (mass airflow, atmospheric temperature, compressor inlet temperature, and steam outlet flow) through correlation analysis. The LSTM-AM model leverages temporal dependencies in time-series data while adaptively weighting influential inputs via attentive learning. Trained with the Adam optimizer to minimize mean squared error, the model achieves superior performance, with a mean absolute percentage error (MAPE) of 0.927 % and a root mean squared error (RMSE) of 4.030 MW, outperforming baseline RNN, GRU, and standalone LSTM models. Furthermore, the model enables proactive load scheduling by predicting hourly demand, reducing unnecessary unit startups/shutdowns, and cutting energy waste. This work contributes a practical framework for power plant operators to optimize dispatch strategies, demonstrating the potential of hybrid deep learning in advancing energy efficiency and operational decision-making.