A Temporal-Guided Graph Multitask Learning Framework for Multiperiod Optimal Power Flow

The data-driven method with strong approximation capabilities can be employed for the efficient computation of multi-period optimal power flow (MPOPF). However, existing methods usually ignore the coupling information among time periods, which significantly increases learning difficulties and the demand for learning samples. To this end, this article proposes a multitask learning framework based on graph neural networks, which integrates the temporal coupling features of MPOPF. This framework enables higher precision of model solving with fewer training samples. Specifically, a temporal graph features construction method based on multivariate correlation entropy is proposed to reduce the learning complexity. Subsequently, a shared layer based on the generalized depth graph convolutional network is designed to extract and share the time period coupling effects for learning ramp constraints, energy storage charging and discharging constraints. Concurrently, a multitask prediction layer based on transformer neural networks has been designed to map each period of MPOPF. This designed architecture achieves both global information sharing and local feature focusing. Moreover, through the joint training approach with multitask adaptive loss weights and feasibility restoration learning strategy, the outputs of the neural network are guaranteed to meet the feasibility. The simulation results show the effectiveness of the proposed methods.