Runoff simulation and forecasting are essential for watershed flood hazard mitigation and optimal utilization of regional water resources. A key factor affecting these processes is the structural heterogeneity of precipitation inputs. As the primary source of precipitation data derived from ground observations, rain gauges can substantially improve rainfall-runoff modeling accuracy when deployed with appropriate density and spatial distribution. This study constructs a spatio-temporal graph neural network framework that integrates Long Short-Term Memory (LSTM) and Graph Neural Network (GNN) approaches to jointly capture the temporal dynamics of hydrological variables and the spatial topological structure among stations for watershed runoff simulation. Meanwhile, the effects of rain gauge density and spatial distribution on model performance are systematically assessed using multiple mean areal precipitation (MAP) estimation methods. The results revealed that: (1) Rain gauge samples selected through clustering form four density distribution scenarios, representing 100%, 72.22%, 50%, and 27.78% of the full network respectively. Across four distributions, the average Nash-Sutcliffe efficiency (NSE) values exceeded 0.93. The runoff simulation associated with Distribution 3 yielded the best performance (NSE=0.967, mean absolute error MAE=175.5m3/s, relative bias BIAS=0.01, and coefficient of determination R2=0.98) under both high-flow and low-flow conditions; (2) Among all the MAP methods, the clustering weight method produced the most robust results, achieving the highest NSE, relatively low MAE, and BIAS closest to 0. Additionally, Distribution 3 maintained the best overall performance, with the smallest simulation errors and biases. The optimization of the rain gauge network combined with the selection of appropriate MAP approaches can enhance the efficiency and adaptability and simulation accuracy of runoff simulation models, offering a solid scientific foundation for hydrological forecasting.