Abstract: The traditional Tolles-Lawson (TL) linear model suffers limitations of compensation precision in processing complex magnetic disturbance data obtained using a towed coherent population trapping (CPT)-based atomic magnetometer detection system. This study introduced neural network technology to achieve more precise nonlinear compensation. Given that conventional neural network models encounter issues such as gradient vanishment or overfitting during parameter updates, this study proposed a deep residual network-based magnetic compensation model (MCRNet) to further enhance the compensation effects of magnetic interference within a towed pod. Experimental results demonstrate that compared to the best traditional method—The Residual Backpropagation Network (ResBP), the proposed MCRNet model increased the improvement ratio (IR) to 4.826 from 4.251, increasing by 13.53%, and reduced the root mean square error (RMSE) of compensated residual magnetism to 0.171 from 0.2, representing a reduction of 14.5%. The proposed model enhances the magnetic survey accuracy of towed magnetic anomaly detection systems.