The weak fault characteristics and the presence of numerous harmonic signals in distribution networks with renewable energy sources reduce the effectiveness of traditional fault diagnosis methods. A fault diagnosis method based on an improved graph neural network was proposed. Wavelet transform was applied to extract the detail coefficients of current and voltage before and after faults. Weighted projection correlation analysis was performed to calculate the correlation between electrical quantities. Highly correlated quantities were selected as inputs to construct the fault diagnosis model using a graph neural network. Fault simulation models for different voltage levels were developed in MATLAB/Simulink. The results indicate that fault signals are effectively enhanced, and faults are accurately located and classified in distribution networks with renewable energy sources at different voltage levels. Good diagnostic performance is maintained in the presence of data loss and noise, demonstrating strong robustness and generalization.