Microseismic source localization is a key component of microseismic monitoring, but conventional methods based on travel-time ray tracing or waveform migration strongly depend on accurate phase picking and reliable subsurface velocity models. Their performance is therefore often degraded in complex engineering environments with heterogeneous geological structures and noisy field data. To address this limitation, this study adapts a classical graph neural network (GNN) model originally developed for natural earthquake localization to gas-storage microseismic monitoring through appropriate parameter configuration and a standardized data-processing workflow. Three-dimensional complex synthetic microseismic data and field data from a gas storage site in Southwest China are used to construct training, validation, and testing datasets. The preprocessing workflow includes abnormal-trace removal, waveform-length normalization, coordinate scaling, and zero-padding for missing station components. The model is implemented and trained using the TensorFlow-Keras framework. Experimental results demonstrate that the adapted GNN achieves feasible and stable localization performance for both synthetic and field datasets. Compared with convolutional neural networks, U-Net, and TransU-Net benchmark models, the GNN provides higher spatial localization accuracy and better stability, with more concentrated source-location results. The proposed approach substantially reduces the dependence on precise phase arrivals and high-resolution velocity models, showing practical robustness under limited data volume and noisy monitoring conditions. This study provides a useful deep learning-based solution for long-term integrity monitoring of gas storage geological bodies and offers a reference for further development of microseismic signal processing and source localization methods.