Efficient fault location of distribution network is the key to ensuring stable and reliable operation of the power grid. Traditional methods mainly focus on static network analysis, often ignoring the dynamic characteristics of faults that evolve over time. To solve this problem, we propose a novel approach that leverages temporal graph convolutional network to enhance fault location accuracy by jointly modeling temporal and spatial correlations. Specifically, we introduce a temporal patch division strategy that processes time-series data in patch units, enabling effective temporal representation. Additionally, a pyramidal feature extraction mechanism is incorporated to capture multiscale temporal dependencies, while attenuation factors are embedded in the graph convolution process to facilitate robust feature integration across time steps. Experimental results on a distribution network fault dataset, generated using the PSCAD simulation platform, demonstrate that the proposed method outperforms existing fault location techniques, significantly improving both accuracy and robustness. The model effectively captures spatiotemporal dependencies, underscoring its potential for fault detection and localization in modern power grids.