Ensemble Entropy with Adaptive Deep Fusion for Short-Term Power Load Forecasting.

Area of Science:

• Electrical Engineering
• Data Science
• Artificial Intelligence

Background:

• Power load forecasting is essential for stable and economical power system operations.
• Traditional methods struggle with the complex, non-stationary nature of power load data.
• Challenges include capturing instantaneous dynamics and fusing multi-feature information effectively.

Purpose of the Study:

• To propose a novel framework, Ensemble Entropy with Adaptive Deep Fusion (EEADF), for short-term multi-feature power load forecasting.
• To improve the accuracy and robustness of power load predictions by addressing limitations of existing methods.

Main Methods:

• Developed an ensemble instantaneous entropy extraction module to compute and fuse approximate, sample, and permutation entropies.
• Implemented a task-adaptive hierarchical fusion mechanism, including feature concatenation and multi-head self-attention fusion.
• Utilized a dual-branch deep learning model processing raw sequences (LSTM) and entropy features (MLP) in parallel.

Main Results:

• The EEADF framework demonstrated robustness in recognizing diverse dynamic patterns on simulated data (MSE: 0.0125, MAE: 0.0794, R²: 0.9932).
• On the real-world ETDataset, EEADF significantly outperformed baseline models (LSTM, TCN, Transformer, Informer) and traditional entropy methods.
• Ablation studies confirmed the importance of entropy features and the fusion mechanism for prediction accuracy.

Conclusions:

• The proposed EEADF framework offers a significant advancement in short-term multi-feature power load forecasting.
• The method effectively captures system dynamics and fuses multimodal information, leading to superior prediction performance.
• EEADF provides a robust and accurate solution for practical power system management.