Partial discharge (PD) surveillance constitutes a pivotal methodology for diagnosing insulation failures in electrical equipment. Enhancing comprehensively the precision of identifying PD anomalies in Gas Insulated Switchgear (GIS) is of paramount significance for ensuring the steady functioning of power grids. This study introduces a novel framework that integrates Phase-Resolved PD Graph Segmentation (PRPD-Grabcut) with a tailored MobileNets-based Convolutional Neural Network (MCNN) to classify GIS-related PD issues. Leveraging image segmentation via PRPD-Grabcut, crucial features are extracted from PRPD diagrams, which then facilitate the construction of the MCNN model. This model employs depth-wise separable convolutions alongside inverted residual architectures to tackle the vanishing gradient dilemma inherent in Deep Convolutional Neural Networks (DCNNs) during GIS PD pattern discernment. Upon the model's subsequent training and validation, empirical evidence illustrates that the PRPD-Grabcut-MCNN hybrid significantly alleviates the computational load and storage requisites of the model, concurrently enhancing the recognition precision and expediting the training process of the neural network. Relative to diverse established lightweight neural network architectures, MCNN manifests superior performance in terms of recognition accuracy, reduced cross-entropy loss, and expedited training duration.

Aiming at the problems of poor prediction effect of non-stationary parameters and single warning rule of UHV converter transformer, this study proposes an intelligent warning method based on decomposition-multi-level cascade network and fuzzy set. Firstly, the integrated empirical modal decomposition technique is used to decompose the target parameter sequence into multiple sub-sequences, and the effective components are screened by the DPR-KLdiv confidence ratio, which is dynamically grouped and reconstructed to form a multilevel feature input; and the multilevel cascade network is constructed by combining multi-device parameters to make the time series prediction. The fuzzy function is further introduced to establish the parameter state mapping rules to expand the alarm triggering conditions. The experiments are validated by actual equipment data, and the local discharge signals of different defects are detected by ultra-high frequency method to enhance the generalization ability of the parameters. The results show that the average RMSE and MAE of this method are 23.21 and 18.47 respectively under the hours step prediction, and the accuracy of the warning is over 90 %, which effectively improves the accuracy of non-smooth parameter prediction and the flexibility of the warning decision.