A Hybrid Neural Network-Based Warning Method for Unbalanced Three-Phase Load in Power Distribution Transformers

This paper proposes a novel hybrid neural network approach for three-phase load imbalance warning in distribution transformers. The method integrates convolutional neural networks (CNN) and bidirectional long short-term memory (BiLSTM) with an attention mechanism to simultaneously capture spatial inter-phase relationships and temporal load evolution patterns. Firstly, phase-aware CNN kernel is employed to explicitly model voltage/current asymmetries. Then, BiLSTM layers are employed to learn long-term load dynamics. Finally, an attention layer is employed for adaptive feature weighting. The experiments on realistic dataset show that the proposed method outperforms baseline methods in terms of warning accuracy. Key advantages include <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{4 0. 5 \%}$</tex> lower false alarm rate than LSTM and maintained robustness under noisy/missing data conditions. This work advances distribution grid monitoring by combining high detection performance with operational practicality.