A New Method for Transformer Fault Prediction Based on Multifeature Enhancement and Refined Long Short-Term Memory
Xin Ma, Hao Hu, Yizi Shang
Abstract
This research proposes a novel predictive model to improve the gas prediction accuracy in transformer oil and provide guarantees for accident prevention. First, this study constructs a cross-entropy loss function with variable thresholds and dynamic weights to reduce error transmission in the deep residual shrinkage network, enhancing the sensitivity of the normal and abnormal transformer states by the network. Second, the multiobjective particle swarm algorithm and random walk strategy are adopted to optimize the long short-term memory (LSTM) network to ensure the prediction model's objectivity. Finally, the improved subchannel threshold depth residual shrinkage network is integrated with the optimized LSTM network. The new model can identify potential abnormal conditions in advance and preproduce an approximate fault model using a feature gas vector (similar to image recognition). Experiments verify the effectiveness of the method. Compared with the existing prediction methods, the new method can avoid blind prediction and significantly improve the prediction accuracy and efficiency, which provides essential value for transformer fault prevention.