A train electromagnetic compatibility risk assessment model based on combined weighting and deep neural networks is proposed to address the complex electromagnetic environment of high-speed railway signaling equipment and the limited capability of conventional pass/fail electromagnetic compatibility (EMC) tests in supporting fault precursor identification. First, a Cube risk evaluation system with threat, vulnerability, and consequence dimensions is established. A three-layer progressive assessment architecture is constructed. Second, subjective and objective weights are fused using analytic hierarchy process (AHP), entropy weight method, and game theory. A static assessment model is built by combining technique for order preference by similarity to ideal solution (TOPSIS) and cloud model for risk ranking and uncertainty characterization. Third, a multilayer perceptron network is designed to output comprehensive risk physical quantities and cloud model features under four operation scenarios, enabling dynamic risk prediction under unknown conditions. Finally, surge immunity tests on axle counter equipment validate the model. The model’s discrimination of the optimal grounding scheme matches the measured results. The proposed model system has good reliability and engineering applicability, providing quantitative decision support for EMC risk management in rail transit.