面向港口能源-物流融合微电网最优调度的多任务可信学习方法

Networked seaport microgrids face the challenge of incurring significant computational costs from solving the joint scheduling problem of energy-logistics integrated microgrids (ELMs) on a daily basis. While existing learning-based optimization approaches offer acceleration, they often fail to generalize across daily variations and struggle to ensure high feasibility. This paper proposes a learning-based method for ELMs' daily scheduling problem with dimension-varying parameters while ensuring high feasibility. To accommodate the varying dimensionality caused by fluctuations in the number of arriving ships, an integer variable predictor is developed by integrating an attention-based neural network with a k-nearest neighbors (KNN) algorithm. The remaining sub-optimization problems are then rapidly solved after these predictions are fixed. A tailored trustworthy repair layer is developed using a feasible region projection approach to rectify infeasible solutions. Experimental results on ELMs data from the Jurong Port in Singapore demonstrate that compared to traditional branch-and-bound and learning-based methods, the proposed method achieves the lowest computation time while maintaining near-optimal cost and 93% feasibility.