Spatio-temporal feature amplified forecasting framework for uncertain power tracking of multitype renewable energy and loads

The integration of multitype renewable energy and loads such as PV, wind power, and EVs, has significantly increased uncertainty in both power supply and demand, which necessitates accurate forecasting to maintain the secure and stable operation of the power grid. However, challenges associated with complex spatio-temporal features hinder the existing forecasting methods from accurately and promptly tracking the instantaneous variations of uncertain power. Therefore, this paper proposes a spatio-temporal feature amplified (STFA) forecasting framework, which can be seamlessly incorporated into state-of-the-art deep learning algorithms. First, a spatio-temporal feature integrated module is constructed that progressively combines phase space reconstruction, positional encoding, and mask. The series of reorganized steps enhance the spatio-temporal features to support training by improving models understanding of uncertain fluctuations. Then, a self-attention mechanism (SAM) is integrated into the deep neural networks (DNNs) to assist training in effectively extracting and utilizing key features. Additionally, a special dynamic weighted loss function with an adaptive bidirectional adjustment mechanism is designed to optimize training by assigning greater importance to abrupt changes. Case studies based on real-world datasets show that the STFA framework accurately tracks fluctuations in uncertain power, especially across multiple prediction targets and DNNs, consistently outperforming methods without the framework.