Research on Distribution Network Fault Prediction Based on Deep Fusion and Salient Feature Extraction from Multimodal Global Data

Electricity serves as a critically important energy source in modern society, with distribution networks constituting key public infrastructure for power transmission. Ensuring the high-quality and stable operation of distribution networks, along with predictive prevention of potential faults, holds significant practical importance. Given the multitude of factors affecting distribution network stability, massive datasets are generated through stability monitoring. A major challenge lies in achieving effective data fusion, interaction, sharing, and feature extraction to synergistically enhance fault prediction capabilities.This paper proposes a Multilayer Perceptron (MLP) fault prediction model for distribution networks, integrating grid-based multimodal global data deep fusion and data-driven salient feature extraction. Accounting for the spatiotemporal characteristics and multi-source heterogeneity of distribution network data, our approach achieves comprehensive data integration through refined grid partitioning. We introduce Random Forest for feature extraction to evaluate inter-feature interactions and nonlinear relationships, thereby identifying salient features. The MLP prediction model is constructed using these extracted features.Experimental results demonstrate that the proposed data fusion and feature extraction methods facilitate thorough data integration and interactive sharing. This provides higher-value feature inputs for the fault prediction model, substantially enhancing both predictive performance and operational reliability.