In the novel power system

the large-scale integration of distributed generation imposes multiple challenges for distribution network fault diagnosis

including complex multi-source information interaction

weakened fault electrical characteristics

and uncertain bidirectional power flows. To address these issues

this paper focuses on fault diagnosis algorithms based on multi-modal feature fusion. Firstly

it employs Variational Mode Decomposition (VMD) to preprocess raw data

reducing noise and non-stationarity

thereby improving the accuracy of the fault diagnosis algorithm. Subsequently

a novel spatio-temporal dual-stream parallel feature extraction network is innovatively designed to mine multi-modal information from both spatial and temporal dimensions. Specifically

a Graph Attention Network (GAT) models the distribution network topology to extract spatial correlation features among nodes

while a Bidirectional Gated Recurrent Unit (BiGRU) captures the temporal evolution patterns of electrical quantities at each node. Furthermore

an adaptive multi-modal feature fusion mechanism is constructed. A self-attention layer dynamically weights and deeply fuses the aforementioned heterogeneous features

thereby generating a more discriminative comprehensive representation for accurate fault classification. The core innovations of this study lie in proposing a novelspatio-temporal dual-stream network architecture and designing an efficient adaptive multi-modal feature fusion strategy

effectively addressing the problems of insufficient fault feature extraction and underutilization of heterogeneous information under high penetration of distributed generation. Experimental results show that the proposed method achieves a significant improvement in fault diagnosis accuracy compared to existing mainstream single-model and hybrid-model diagnostic approaches

demonstrating its superiority

robustness

and strong generalization capability in complex new power system scenarios.