Optimization of Transparency of Operational Risks in Substations Based on Big Data Analytics

Substations are important nodes in the power grid, and their operational safety directly affects the reliability of power supply and the lives and property of the people. However, there are difficulties in integrating multi-source heterogeneous data, existing intelligent analysis methods have "black box" characteristics, and traditional risk management methods are difficult to apply to dynamic coupling and other problems. This paper investigates a novel method for optimizing the transparency of substation operational risks based on a time-space graph attention network, aiming to transition from "black box prediction" to "white box explanation." The specific research approach is as follows: First, a heterogeneous dynamic graph model of "human-machine-environment" is established, mapping four types of data—ultra-wideband positioning, video surveillance, environmental perception, and SCADA—to a unified 128-dimensional space, and a sliding window mechanism is used to generate time-series sequences. Second, a time-space graph attention network, ST-GAT, is designed, using an 8-head heterogeneous graph attention mechanism to capture spatial coupling relationships and dilated causal convolution to uncover temporal evolution patterns. Risk attribution explanation is achieved through attention weights. Finally, a risk situation-driven multi-objective optimization model is established, employing the NSGA-II algorithm to solve for the Pareto optimal scheduling scheme, simultaneously optimizing the three objectives of safety, efficiency, and transparency. The experiment was conducted based on real data from a 220kV substation. Results show that the ST-GAT risk identification achieved an F1-score of 0.891 and an AUC-ROC of 0.931, a 4.6% improvement over ST-GCN. Its explained coverage was 87.3%, and its manual review consistency was 0.84, significantly better than GAT’s 62.5% and 0.61, respectively. The ST-GAT-NSGA-II optimization framework achieved a 42.7% risk reduction, an 18.5% efficiency improvement, and a decision acceptance rate of 0.89. Its hypervolume index of 0.467 was a 31.2% improvement over the weighted summation method. The proposed method effectively fills the application gap of graph neural networks in the field of power operation safety, providing an interpretable technical paradigm for intelligent safety management and control of substations.