Dynamic load-balancing optimisation with bidirectional edge detection under multi-scale feature fusion

This paper proposes a scheduling optimisation framework GCN-EDQNet that integrates multi-scale feature fusion, bidirectional edge detection, and graph convolutional network (GCN).Data centre resources are modelled as a graph over distributed nodes; a multi-scale module builds hierarchical representations to capture spatial heterogeneity in resource distributions.A bidirectional edge-detection subnetwork then identifies scheduling-sensitive regions and produces an edge heatmap -assigning higher weights to edges connecting nodes with high load variation -that guides the GCN to prioritise structurally salient, mutation-prone areas.This explicit weighting mechanism enables the GCN to focus on bottleneck-prone regions and improve structure-aware feature learning.Finally, a reinforcement learning strategy enables adaptive task allocation and migration.Experiments on two public datasets show that GCN-EDQNet outperforms conventional approaches in task completion time, load variance, scheduling success rate, and energy efficiency.These results highlight a structure-aware, intelligent paradigm for data centre resource scheduling with clear theoretical and practical value.