Traditional power grid O&M systems suffer from two limitations related to their architecture. First, fault diagnostic models consider equipment units individually, ignoring dependencies between devices that are implicit in grid topology. Second, there is no clear link between diagnostic outcomes and maintenance recommendations, making the latter process depend heavily on ad-hoc decision making. The current work presents a design approach that seeks to address both limitations with two particular architectural solutions. The first solution introduces a representation of the state of the equipment, based on a combination of temporal features extracted using deep learning and neighborhood context computed using a graph attention network. This allows representing dependencies between faults without explicitly designing a new fault diagnosis model. The second innovation is a bridging component that translates diagnosis results into constraint-based maintenance recommendations in the form of a ranked list of recommendations for O&M personnel. The design was evaluated on a real-world dataset of 12,847 annotated examples across 23 classes of faults at six substations. The achieved accuracy of fault diagnosis is 94.3 ± 0.4% (p<0.01) and NDCG@5 of 0.887 ± 0.012, which constitute 2.9 and 10.4-percentage-point improvements over the baselines.
