Learning picturized and time-series data for fault location with renewable energy sources

Transmission lines are heavy assets of power systems. Therefore, the rapid and accurate identification of fault locations is important for power grids' safe and stable operation. Traditional methods based on impedances or traveling waves are facing increasing challenges like uncertain power generation and unknown power electronic device characteristics when modern power systems are transitioning to deeply renewable energy source (RES) penetrated grids. In order to solve the above problems, this paper proposes a high-dimensional time-frequency feature extraction method that does not require expert knowledge of physical features. This paper proposes a fault location framework for learning picturized and time-series data (FltLoc-FPTD) with renewables. The developed loss function is suitable for the classification of faulty lines, considering the challenges of distinguishing faults in adjacent lines. Furthermore, we design an enhanced convolutional neural networks (CNN) subsampling layer with blur kernels to replace the traditional subsampling layer to eliminate the influence of high-frequency noise and improve the robustness against noise. The effectiveness of the method is verified by simulation under two benchmark systems. The average fault location errors with and without environment noises are 0.0189 and 0.0124.