DataSheet1_Multi-task learning load time series situational prediction based on gated recurrent neural networks considering spatial correlations.docx

Accurate load forecasting plays a crucial role in the effective planning, operation, and management of modern power systems. In this study, a novel approach to load time series situational prediction is proposed, which integrates spatial correlations of heterogeneous load resources through the application of Random Matrix Theory (RMT) with a Multi-Task Learning (MTL) framework based on Gated Recurrent Units (GRU). RMT is utilized to capture the complex, high-dimensional statistical relationships among various load profiles, enabling a deeper understanding of the underlying data patterns that traditional methods may overlook. The GRU-based MTL framework is employed to exploit these spatiotemporal correlations, allowing for the sharing of essential features across multiple tasks, which in turn enhances the accuracy and robustness of load predictions. This approach was validated using real-world data, demonstrating notable improvements in prediction accuracy when compared to single-task learning models. The results indicate that this method effectively captures complex relationships within the data, leading to more accurate load forecasting. This enhanced predictive capability is expected to contribute significantly to improving demand-side management, reducing the risks of grid overloading, and supporting the integration of renewable energy sources, thereby fostering the overall sustainability and resilience of power systems.