Contextually enhanced ES-dRNN with dynamic attention for short-term load forecasting.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Time Series Analysis

Background:

• Accurate short-term load forecasting (STLF) is crucial for power grid management.
• Existing models often struggle with complex dependencies and dynamic data variations.

Purpose of the Study:

• To develop a novel STLF model with enhanced accuracy and robustness.
• To leverage hybrid architectures and contextual information for improved forecasting.

Main Methods:

• A hybrid and hierarchical architecture combining Exponential Smoothing (ES) and Recurrent Neural Network (RNN).
• A dual-track system (context and main) for dynamic information integration.
• Attentive dilated recurrent cells to capture multi-scale temporal dependencies.

Main Results:

• The proposed model significantly outperforms its predecessor and standard statistical/ML models.
• Demonstrated superior accuracy across 35 diverse forecasting problems.
• Successfully captures short-term, long-term, and seasonal patterns.

Conclusions:

• The contextually enhanced hybrid model offers a significant advancement in STLF.
• The architecture effectively integrates diverse temporal information for precise load predictions.
• The model provides both point forecasts and reliable predictive intervals.