GridRiskNet: A Two-Stage Hybrid Model for Project Investment Risk Management of Power Grid Enterprises Using Big Data Mining

Abstract

To enhance the power grid enterprise's ability to comprehensively perceive and dynamically assess investment risks in engineering projects, this study proposes a risk management model called GridRiskNet based on big data mining. This model integrates structured, unstructured, and spatiotemporal data and realizes intelligent identification of project risk probability distributions and potential impact ranges by constructing a two-stage hybrid modeling architecture. In the first stage, the model uses eXtreme Gradient Boosting (XGBoost) and Light Gradient Boosting Machine (LightGBM) to extract static and dynamic features in parallel. In the second stage, it introduces Graph Attention Recurrent Neural Network (GA-RNN) to model risk propagation paths under the power grid topology. Meanwhile, this study combines Spatio-Temporal Graph Convolutional Network (ST-GCN) to improve the coupling expression of meteorological and text features. The experiment uses multi-source public data for verification, such as power infrastructure data from the U.S. Energy Information Administration, meteorological observation data from the National Oceanic and Atmospheric Administration, and power grid topology data from OpenStreetMap. The results show that GridRiskNet performs excellently in risk prediction stability and regional propagation modeling. Among them, the risk principal component analysis projection score in 2023 reached 7.779. This indicates that cost overruns, climate pressure, and equipment technology risks together form a high-risk cluster, with cost overruns increasing by 269% compared with 2018. In the State-of-the-Art comparison, GridRiskNet achieves an F1-score of 0.892, a Receiver Operating Characteristic - Area Under Curve of 0.962, a Risk Impact Radius error of approximately 4.8 km, and a Risk Entropy of 0.89; these are comprehensively better than existing methods. Moreover, the model has good cross-modal feature fusion and risk transmission mechanism identification capabilities, and can effectively characterize the spatiotemporal coupling risk features in complex power grid projects. Overall, this system can provide power grid enterprises with structured and interpretable risk index outputs and regional early warning support. Thus, it helps to improve the investment safety and operational and maintenance resilience of projects.