The power supply quality and power supply safety of a low-voltage residential power distribution system is seriously affected by the occurrence of series arc faults. It is difficult to detect and extinguish them due to the characteristics of small current, high stochasticity, and strong concealment. In order to improve the overall safety of residential distribution systems, a novel method based on discrete wavelet transform (DWT) and deep neural network (DNN) is proposed to detect series arc faults in this paper. An experimental bed is built to obtain current signals under two states, normal and arcing. The collected signals are discomposed in different scales applying the DWT. The wavelet coefficient sequences are used for forming training set and test set. The deep neural network trained by training set under 4 different loads adaptively learn the feature of arc faults. The accuracy of arc faults recognition is sent through feeding test set into the model, about 97.75%. The experimental result shows that this method has good accuracy and generality under different types of loading.

Accurate prediction of hydraulic support pressure is crucial for ensuring coal mine safety. With increasing mining depths and increasingly complex operating environments, precise prediction faces greater challenges. To address these challenges, this study proposes an LSTM-PatchTST prediction method based on multi-dimensional feature dependency fusion: First, Pearson correlation analysis is used to screen key features, and Gaussian moving average filtering optimizes the data; Subsequently, the preprocessed time series is input to the LSTM network, where the forget gate and input gate capture short-term fluctuations and long-term trends respectively, while residual connections ensure complete preservation of multi-layer temporal features; Then, the dynamic features extracted by LSTM are passed to the PatchTST module, which divides the sequence into local patches, with a multi-layer self-attention encoder simultaneously modeling local details and global dependencies, achieving deep feature fusion. The model’s performance is validated using actual pressure data from Fucun Coal Mine in Zaozhuang, Shandong. Experimental results show that, compared to pure PatchTST and Transformer + LSTM models, the proposed model reduces RMSE by approximately 48.6% and 30.0%, and MAE by approximately 58.7% and 38.8%, respectively. Finally, to further verify the model’s generalization ability, the trained model was transferred to a dataset from Gengcun Coal Mine in Yima, Henan, where compared to pure PatchTST and Transformer + LSTM models, the proposed model reduces RMSE by approximately 34.6% and 31.4%, and MAE by approximately 35.7% and 29.9%, respectively.

For low-voltage three-phase systems, the deep fault arc features are difficult to extract, and the phase information has strong timing. This phenomenon leads to the problem of low accuracy of fault phase selection. This paper proposes a three-phase fault arc phase selection method based on a global temporal convolutional network. First, this method builds a low-voltage three-phase arc fault data acquisition platform and establishes a dataset. Second, the experimental data were decomposed by variational mode decomposition and analyzed in the time-frequency domain. The decomposed data are reconstructed and used as input to the model. Finally, in order to reduce the fault features lost during the causal convolution operation, the global attention mechanism is used to extract deep fault characterization to identify faults and their differences. The experimental results show that the accuracy of the three-phase arc fault arc phase selection of the model can reach 98.62%, and the accuracy of single-phase fault detection can reach 99.39%. This model can effectively extract three-phase arc fault and phase characteristics. This paper provides a new idea for series fault arc detection and three-phase fault arc phase selection research.

The operational dynamics of residential appliances impede arc fault identification, particularly in multi-branch circuits where fault signatures are obscured by load interference, increasing diagnostic complexity. To address this challenge, we develop the CASCNet diagnostic framework through synergistic integration of ShuffleNetV2's hierarchical feature learning and CatBoost's categorical processing. Firstly, synchronous compression wavelet transform converts arc signals into 2D time-frequency images to enhance clarity and reduce noise. Next, ShuffleNetV2 efficiently extracts features, optimized with a CA attention mechanism to reduce complexity and improve channel interaction. Finally, CatBoost’s robust generalization and noise resistance classify features. Experiments show this method accurately identifies series arc faults and load types, maintaining strong performance under noise, offering an effective solution.