An Attention-Based Multilayer GRU Model for Multistep-Ahead Short-Term Load Forecasting^†.

Area of Science:

• Artificial Intelligence
• Electrical Engineering
• Data Science

Background:

• Multistep-ahead prediction is crucial for electric load forecasting, especially for handling sudden power consumption changes.
• Recurrent Neural Networks (RNNs), like Gated Recurrent Units (GRUs), are effective for time-series prediction due to their ability to learn from past data.
• Standard GRU models have limitations in prediction accuracy as they treat all input variables equally.

Purpose of the Study:

• To develop an advanced short-term load forecasting model.
• To enhance the prediction accuracy of GRU networks by enabling them to prioritize important input variables.
• To improve multistep-ahead forecasting performance, particularly for long input sequences.

Main Methods:

• Proposed an attention-based Gated Recurrent Unit (GRU) model for electric load forecasting.
• Implemented a mechanism within the GRU to assign differential importance to input variables.
• Conducted extensive experiments to evaluate the model's performance on building-level power consumption data.

Main Results:

• The attention-based GRU model demonstrated significant performance improvements in multistep-ahead prediction.
• The model showed enhanced accuracy, especially when dealing with longer input sequences.
• Outperformed other recent multistep-ahead prediction models in building-level load forecasting tasks.

Conclusions:

• The proposed attention-based GRU model offers a superior approach to short-term electric load forecasting.
• Focusing on crucial variables via attention mechanisms significantly boosts prediction accuracy.
• This model represents a notable advancement for accurate building-level power consumption forecasting.