Data-driven fault location method for multi-terminal power network with physics-AI mapping

The scale of modern power networks is growing every year. An accurate fault location method is necessary to ensure the reliability of the power supply. Existing data-driven methods typically fail to consider the physics-AI mapping to ensure the physical meaning of the trained network. In addition, for a power network, it is hard for the existing methods to pick the proper measurements used for fault location. This paper proposes a data-driven method that can decide the data from which measurement should be focused on and can identify the fault location by itself. The proposed method is based on multiple voltage measurements at terminals of the network. First, continuous wavelet transform is applied to extract time and frequency features within the waveforms. Afterwards, the extracted features are fed into convolutional neural network with self-attention mechanism. Based on self-attention mechanism, the proposed method can adjust the weights of a data driven network for different inputs. As a result, even when data from one measurement was lost, it can still use the rest of the measurements to identify the fault location. Finally, Grad-CAM is used as a tool to demonstrate the physical-AI mapping, showing how different physical features within each measurement are considered. The results clearly show the effectiveness of this method.