With the increasing complexity of FPGA design, physical design requires a large number of optimization iterations to achieve. Cabling congestion affects chip area, delay and other performance indicators, so accurate and rapid prediction and early resolution are required. A FPGA routing congestion prediction model CBAM-CGAN is proposed. The model extracts feature in the layout phase to synthesize learning images, and introduces attention mechanism learning to enhance the importance of each feature channel of the image, so as to improve the routing congestion prediction performance. The experimental results show that the method achieves good results in routing congestion prediction in the layout phase. Compared with the conditional countermeasure generation network model, the average value of structure similarity is increased by 0.89%, the average value of peak signal to noise ratio is increased by 1.37%, the average value of normalized root mean square pixel difference is decreased by 3.8%, the average value of pixel accuracy difference is decreased by 0.06%, and the prediction time of a single image is about 0.1 seconds. Experimental data prove the accuracy and rapidity of the model in FPGA routing congestion.