Power system load forecasting using mobility optimization and multi-task learning in COVID-19.

Area of Science:

• Energy Systems Analysis
• Artificial Intelligence in Energy
• Pandemic Impact Studies

Background:

• The COVID-19 pandemic significantly altered global energy production and consumption patterns.
• Conventional short-term load forecasting methods may struggle to accurately capture these new load dynamics.
• The effectiveness of existing models under pandemic conditions requires thorough investigation.

Purpose of the Study:

• To develop and validate an advanced load forecasting method resilient to pandemic-induced disruptions.
• To assess the limitations of traditional forecasting approaches during the COVID-19 era.
• To enhance the accuracy and robustness of short-term load predictions.

Main Methods:

• Proposed a novel mobility-optimized load forecasting method integrating multi-task learning and long short-term memory (LSTM) networks.
• Incorporated real-world mobility data and data-sharing layers to improve pattern recognition and model generalization.
• Utilized Shapley Additive Explanations (SHAP) for objective model interpretation and validation.

Main Results:

• Conventional forecasting models demonstrated failure in accurately predicting loads during the COVID-19 pandemic.
• The proposed mobility-optimized LSTM model achieved high accuracy, with prediction errors consistently below 1%.
• SHAP analysis confirmed the significant contribution of mobility indicators to accurate load forecasting and revealed non-synchronous indicator-load relationships.

Conclusions:

• The developed multi-task learning approach effectively addresses the challenges posed by the COVID-19 pandemic to short-term load forecasting.
• Mobility data is a crucial indicator for improving the accuracy and robustness of energy load predictions.
• The study highlights the need for adaptive and data-driven forecasting models in response to unforeseen global events.