Scholarlink

Accurate multi-energy load forecasting is essential for the stable operation and optimal dispatch of integrated energy systems (IES). This study introduces a forecasting model combining time series decomposition and deep learning to tackle challenges from volatile and stochastic loads. Firstly, Meteorological factors strongly correlated with load variation are chosen as input features. Secondly, the complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) method is applied to decompose highly fluctuating heating load into intrinsic mode functions (IMFs), reducing noise interference. Then, a 3D convolutional neural network is reconstructed into a three-layer 1D convolution to enhance feature extraction and generalization. Finally, an attention mechanism is introduced before the output layer to enable differential extraction of key features from the shared layer for each subtask. The results show that the proposed model achieves MAPE as low as 1.183%, 1.137%, and 1.578% across different seasons, significantly reducing error accumulation and delivering the highest accuracy.

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 Integrated Energy Systems Multi-energy Load Deep Learning Joint Prediction Method Based on CEEMDAN

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Guangshuo Liu,
  • Ke Li,
  • Yuchen Mu,
  • Chenghui Zhang

摘要

Accurate multi-energy load forecasting is essential for the stable operation and optimal dispatch of integrated energy systems (IES). This study introduces a forecasting model combining time series decomposition and deep learning to tackle challenges from volatile and stochastic loads. Firstly, Meteorological factors strongly correlated with load variation are chosen as input features. Secondly, the complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) method is applied to decompose highly fluctuating heating load into intrinsic mode functions (IMFs), reducing noise interference. Then, a 3D convolutional neural network is reconstructed into a three-layer 1D convolution to enhance feature extraction and generalization. Finally, an attention mechanism is introduced before the output layer to enable differential extraction of key features from the shared layer for each subtask. The results show that the proposed model achieves MAPE as low as 1.183%, 1.137%, and 1.578% across different seasons, significantly reducing error accumulation and delivering the highest accuracy.