The new types of transformer oil currently in use contain dissolved gases, resulting in complex associated fault data. Existing intelligent detection methods, which primarily rely on neural networks, often struggle to locate the global optimal solution in complex temporal environments, thereby compromising monitoring effectiveness. This study proposes a transformer fault diagnosis method for power systems using a genetic algorithm-optimized dense network and dissolved gas analysis. A hybrid convolutional and long short-term memory neural network is constructed, which takes the dissolved gas in transformer oil from the novel power system as the input. Based on this input, convolutional neural networks are first employed to extract spatial features of the fault information, followed by the use of recurrent neural networks with long short-term memory to identify temporal patterns. To enhance the performance of the hybrid network and ensure the accuracy of the output results, the optimal coefficients and offsets of the network are determined through an adaptive genetic algorithm. Finally, the improved hybrid network achieves automatic monitoring of transformer faults. The results indicate that this method can effectively monitor transformer high- and low-temperature overheating faults, as well as low-energy discharge faults. The missed detection rate and false alarm rate are both below 2% (with 95% confidence intervals of 1.4%, 2.1% and 1.2%, 2.3%, respectively), and the Kappa coefficient reaches 0.96. These results are validated through tenfold cross-validation and robustness tests, demonstrating the model’s consistency and reliability under varying data conditions.
Zhang et al. (Fri,) studied this question.