Obtaining vertical ground reaction force (vGRF) data from stroke patients is highly challenging, and analyzing and predicting vGRF and joint moment in stroke patients has high value in sports biomechanics. Objective: Use principal component analysis (PCA) and BP neural networks to predict the joint torque of the affected side of the hip, knee, and ankle in stroke patients during walking. Method: Thirty stroke patients（age 61.5±3.5 years old, weight 68.2±2.4 kg, height 167.2±3.4 cm）were synchronously collected kinematic and dynamic data using an 8-lens Qualisys infrared point high-speed motion capture system and a Kistler three-dimensional force measurement platform. Calculate hip knee ankle joint torque in stroke patients using Opensim, screen initial variables with a cumulative contribution rate of 99% using PCA, and use normalized root mean square error (NRMSE), root mean square error (RMSE), mean absolute percentage error (MAPE), and mean absolute error (MAE) R2 serves as an evaluation indicator for the PCA-BP model. Evaluate the consistency between calculated joint torque and predicted torque using Kendall's W coefficient. Result: PCA data shows that the trunk, pelvis, affected hip joint, knee joint, and ankle joint have a significant impact on the torque of the affected hip, knee, and ankle joint on the x, y, and z axes (sagittal, coronal, and vertical axes, respectively)..The NRMSE between predicted and measured values is 5.14%~8.86%, RMSE is 0.184~0.371, MAPE is 3.5-4%, MAE is 0.143~0.248, and R2 is 0.998~0.999. Conclusion: The established PCA-BP model can accurately predict the hip knee ankle joint torque of stroke patients during walking, significantly shortening the measurement time. This model can replace traditional joint torque calculation in gait analysis of stroke patients, providing a new approach to obtain biomechanical data of stroke patients and an effective method for clinical treatment of stroke patients.