MLFGCN: short-term residential load forecasting via graph attention temporal convolution network.

Area of Science:

• Energy Systems
• Artificial Intelligence
• Data Science

Background:

• Residential load forecasting is complex due to inherent fluctuations and individual variations.
• Current models often rely on external factors like climate, increasing computational load and uncertainty.
• A need exists for accurate forecasting models that minimize reliance on external data.

Purpose of the Study:

• To propose a novel Multi-Level Feature Fusion model (MLFGCN) for enhanced short-term residential load forecasting.
• To develop a model that captures long-term dependencies and inter-series correlations without external information.
• To improve forecasting accuracy and reduce computational burden compared to existing methods.

Main Methods:

• Utilized a Temporal Convolutional Network (TCN) with a gating mechanism to learn long-term dependencies within load series.
• Designed two graph attentive convolutional modules to capture multi-level dependencies in load data.
• Implemented an information fusion layer to integrate outputs from different modules for final forecasting.

Main Results:

• The MLFGCN model demonstrated superior performance on two real-world datasets.
• Achieved Mean Absolute Error (MAE) of 0.25, Mean Absolute Percentage Error (MAPE) of 7.58%, and Root Mean Square Error (RMSE) of 0.50.
• Significantly outperformed baseline models in forecasting accuracy.

Conclusions:

• The MLFGCN algorithm significantly improves short-term residential load forecasting accuracy.
• The model excels through high-quality feature reconstruction, comprehensive graph construction, and spatiotemporal feature capture.
• MLFGCN offers a robust solution for accurate load prediction without external data dependencies.