Multi-Granularity Autoformer for long-term deterministic and probabilistic power load forecasting.

Area of Science:

• Electrical Engineering
• Artificial Intelligence
• Time Series Analysis

Background:

• Long-term power load forecasting is essential for power system planning.
• Existing transformer models face challenges with complexity and parameter overhead.
• Intricate temporal patterns in load data pose forecasting difficulties.

Purpose of the Study:

• Introduce a novel Multi-Granularity Autoformer (MG-Autoformer) for efficient long-term load forecasting.
• Improve the accuracy of predicting future electricity demand.
• Enable probabilistic forecasting and uncertainty quantification.

Main Methods:

• Developed a Multi-Granularity Auto-Correlation Attention Mechanism (MG-ACAM) to capture dependencies at various granularities.
• Implemented a shared query-key (Q-K) mechanism for efficiency and complexity reduction.
• Incorporated a quantile loss function for probabilistic predictions.

Main Results:

• MG-Autoformer demonstrated superior performance in long-term load point forecasting.
• The model achieved excellent results in probabilistic forecasting tasks.
• Experiments confirmed effectiveness on diverse international datasets.

Conclusions:

• The proposed MG-Autoformer effectively models complex temporal dependencies for accurate long-term load forecasting.
• The model offers an efficient and robust solution for both point and probabilistic forecasting.
• MG-Autoformer advances the state-of-the-art in power load prediction.