Adaptive Fusion Method of Multi-Source Fault Data of Power System Based on Machine Learning Technology

The heterogeneity, temporal dynamics, and topological correlation of multi-source fault data in smart grids pose challenges to the accuracy and real-time performance of fault diagnosis. Traditional static modeling methods lack adaptability in dynamic scenarios, and machine learning model feature fusion strategies lack flexibility. This article proposes an adaptive fusion method based on machine learning, which constructs a unified modeling framework for multi-source data through the snowflake model, achieving spatiotemporal alignment between structured electrical measurement and unstructured data; Design a cascaded structure of Dynamic Graph Convolutional Network (DGCN) and Long Short Term Memory Network (LSTM) to decouple the topological associations and temporal dependencies of data; Introducing an adaptive weight optimization mechanism driven by reinforcement learning (RL), dynamically adjusting the fusion strategy based on data reliability; Adopting cloud edge collaborative architecture to improve system response speed. The experiment shows that this method has a fault diagnosis accuracy of 98.7 % on the standard test set, with a processing delay as low as 28 ms. In the scenario of noise intensity <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\sigma=0.3$</tex>, the accuracy remains at 89.1 %, and the response time after topology change is only 50 ms, which is significantly better than traditional weighted fusion, static GNN and other methods. The research results provide a new paradigm for dynamic modeling and intelligent decision-making in the diagnosis of complex faults in smart grids. In the future, we can further explore the integration of physical priors in transfer learning, model lightweight design, and cross regional federated learning.