A Multi-Task Joint Learning Framework for Transformer Fault Prediction

Transmission equipment failure in the power system can disrupt electricity supply and affect normal operations. The current approach for transformer fault prediction is to first predict and then diagnose the fault, but this approach may introduce accumulative errors in prediction and diagnosis tasks. To address these issues, a multi-task learning framework is proposed to optimize forecasting and diagnosis tasks simultaneously. Specifically, based on the history values, we extract the temporal features firstly. After that, two branches are added behind the model to be optimized based on the forecasting values and diagnosis categories respectively. By doing so, we break the ‘forecasting-then-diagnosis’ paradigm and couples the training procedure between forecasting and diagnosis, leading to more accurate fault prediction. Extensive experiments are conducted on practical datasets collected from a 500kV substation. Our framework achieves more accurate fault prediction, which demonstrates the effectiveness of proposed framework.