BILSTM-SimAM: An improved algorithm for short-term electric load forecasting based on multi-feature.

Area of Science:

• Electrical Engineering
• Artificial Intelligence
• Signal Processing

Background:

• Growing user-side resources in distribution systems cause imbalances, necessitating accurate short-term power load forecasting.
• Existing forecasting methods struggle with multi-feature extraction and noise reduction in load data.

Purpose of the Study:

• To develop an advanced deep learning model for improved electric load forecasting.
• To enhance the extraction of multi-features from load data and emphasize key historical information.

Main Methods:

• Variational Mode Decomposition (VMD) to denoise and decompose load data into Intrinsic Mode Functions (IMF).
• Convolutional Neural Network (CNN) with Dropout for improved feature recognition and faster convergence.
• Bidirectional Long Short-Term Memory (BILSTM) combined with a parameter-free attention mechanism (SimAM) for multi-feature extraction.

Main Results:

• The BILSTM-SimAM model achieved an R² of 97.8%, exceeding Transformer, MLP, and Prophet models.
• Demonstrated superior performance with reduced error metrics compared to mainstream forecasting models.
• Validated the effectiveness of VMD, CNN, and BILSTM-SimAM integration for load forecasting.

Conclusions:

• The proposed BILSTM-SimAM network offers a robust and accurate solution for short-term power load forecasting.
• The novel approach effectively handles data noise and extracts critical features for enhanced prediction.
• This method presents a significant advancement in smart grid load management and stability.