Deep residual networks with convolutional feature extraction for short-term load forecasting.

Area of Science:

• Artificial Intelligence
• Electrical Engineering
• Data Science

Background:

• Deep learning models face challenges in balancing feature extraction and temporal representation for Short-Term Load Forecasting (STLF).
• Existing methods often lack generalization across diverse climatic conditions, impacting forecasting accuracy.

Purpose of the Study:

• To develop a Convolutional Neural Network-Embedded Deep Residual Network (CNN-Embedded DRN) for improved STLF.
• To enhance local feature extraction and temporal pattern recognition using CNNs within a DRN framework.
• To evaluate the model's generalization and robustness across different climate zones.

Main Methods:

• Integration of Convolutional Neural Network (CNN) for local feature extraction into a Deep Residual Network (DRN).
• Application of residual learning to improve network stability and mitigate gradient degradation.
• Comparative performance evaluation against baseline and ablation models on temperate (ISO-NE) and tropical (Malaysia) datasets.
• Validation of statistical significance and seasonal robustness using bootstrap analysis.

Main Results:

• The CNN-Embedded DRN achieved the lowest Mean Absolute Percentage Error (MAPE), with 1.5303% on ISO-NE and 5.0566% on Malaysia datasets.
• The model demonstrated superior predictive performance compared to all baseline and ablation models.
• Bootstrap analysis confirmed the statistical significance of the performance improvements.

Conclusions:

• The proposed CNN-Embedded DRN offers a reliable and generalizable framework for STLF.
• The model exhibits improved accuracy, robustness, and adaptability to varying climatic and demand conditions.
• Future work includes extending the framework for multi-regional forecasting and incorporating attention mechanisms.