A Novel Sequence-to-Sequence-Based Deep Learning Model for Multistep Load Forecasting.

Area of Science:

• Electrical Engineering
• Computer Science
• Artificial Intelligence

Background:

• Accurate load forecasting is essential for efficient power system operation, including balancing supply and demand and reducing energy costs.
• Existing methods often focus on single-step forecasting, limiting their utility for optimizing resource allocation and decision-making.
• Multistep load forecasting offers enhanced insights for strategic energy management.

Purpose of the Study:

• To propose a novel deep learning model for accurate multistep load forecasting.
• To leverage time series decomposition within a sequence-to-sequence framework.
• To improve upon existing load forecasting techniques for better energy management.

Main Methods:

• A sequence-to-sequence (Seq2Seq) deep learning model incorporating a time series decomposition strategy.
• The model comprises residual-connected basic blocks, each with a Temporal Convolution Network (TCN) encoder and Long Short-Term Memory (LSTM) decoders.
• Individual forecasting within each basic block, with final results aggregated.

Main Results:

• The proposed Seq2Seq model demonstrated superior accuracy in multistep load forecasting.
• Evaluated on multiple real-world datasets, the model outperformed several benchmark forecasting methods.
• The TCN-LSTM architecture effectively captures temporal dependencies for precise predictions.

Conclusions:

• The novel Seq2Seq deep learning model with time series decomposition provides a significant advancement in multistep load forecasting.
• This approach offers improved accuracy and insights for energy resource allocation and decision-making in power systems.
• The model's performance validates its effectiveness against established forecasting techniques.