Data-driven transmission line fault location with data-efficient transfer learning

Transmission line fault location is one of the essential steps to ensure power supply reliability. Traditional model based methods and traveling wave based methods have limitations such as requirements of accurate line parameters or high sampling rates. Existing data-driven methods usually require large number of training data that are exactly consistent with the practical power system. However, the number of fault data in practical systems are usually quite limited, and there could be mismatch between the practical system and the simulation system, limiting the fault location accuracy. To this end, this paper proposes a transfer learning based data-driven fault location method for transmission lines. The method can efficiently utilize small dataset in practical power systems. First, a neural network is constructed and is pre-trained with extensive data generated by the simulation system A. Next, another very small dataset is generated by simulation system B to mimic the practical scenario, where the line parameters are different from simulation system A. The transfer learning efficiently utilizes the small dataset to update the neural network, with the steps of freeze-training and fine-tuning. Finally, the performances of data-driven methods with and without transfer learning are compared. The results clearly indicate the effectiveness and necessity of the proposed transfer learning based fault location method.