To address the complex spatial-nonlinear relationships between mountainous road structural features and traffic risks, this study proposes a TransRisk Geographically Weighted Neural Network (TGWNN) modeling framework. By designing a Geographically Weighted Layer (GWL), this method deeply integrates spatial location encoding with neural networks and adaptively learns local spatial correlations through a Geographical Weight Adjuster (GWA), thereby simultaneously capturing spatial heterogeneity and high-order nonlinear coupling effects. Simulation-based experiments demonstrate that TGWNN achieves significantly superior prediction performance under strong nonlinearity/spatial heterogeneity scenarios (RMSE = 0.146), outperforming traditional Geographically Weighted Regression (GWR, RMSE = 0.180) and Artificial Neural Networks (ANN, RMSE = 0.798). In empirical applications on mountain roads in Ningxia, TGWNN reduced quarterly prediction RMSE (0.659–1.102) by 31.7% on average compared to baseline models, while successfully visualizing spatial response patterns of structural indicators such as bridge-tunnel ratios and longitudinal gradients. This research confirms TGWNN’s dual capabilities in nonlinear representation and spatial adaptability, providing methodological support for high-precision mountain traffic risk early-warning systems.