DG-LSTM-SA model: A deep gated LSTM network with self-attention mechanism for power generation and load forecasting

Area of Science:

• Artificial Intelligence
• Machine Learning
• Energy Systems

Background:

• Accurate power forecasting is crucial for reliable energy system operation.
• Recurrent Neural Networks (RNNs) struggle with long-term dependencies, while Transformers are computationally intensive.
• Existing models face challenges in balancing accuracy and efficiency for complex power time series.

Purpose of the Study:

• To introduce a novel Deep Gated Long Short-Term Memory network with Self-Attention (DG-LSTM-SA).
• To enhance the accuracy and computational efficiency of power generation and load demand forecasting.
• To address the limitations of traditional RNNs and Transformer-based models in energy time series analysis.

Main Methods:

• Developed a Deep Gated Long Short-Term Memory network with Self-Attention (DG-LSTM-SA).
• Integrated a multi-layer gated architecture with hierarchically embedded self-attention modules.
• Evaluated the model on three real-world energy datasets: NEPOOL, Yichang, and Solar-Energy.

Main Results:

• DG-LSTM-SA consistently outperformed ten baseline models, including LSTM and GRU variants.
• Achieved a reduction in Mean Absolute Error by over 75% compared to standard RNNs.
• Demonstrated competitive accuracy with state-of-the-art attention models while offering superior computational efficiency and faster training.

Conclusions:

• The proposed DG-LSTM-SA model is robust, accurate, and practical for real-world applications.
• DG-LSTM-SA effectively captures complex temporal patterns and long-term dependencies in energy data.
• The model offers a significant improvement for grid dispatch and operational decision-making in energy systems.