The error modeling and experimental analysis were discussed for a semi-symmetrical 3-translational Delta-CU parallel mechanism, which was proposed by the author's team. On the basis of the planning executive terminal trajectory, kinematics error of the actuator was compensated by adopting external direct calibration and correcting the system input. In the process of direct external calibration, the global least square method was used to solve the coordinate transformation parameters，which could reduce the impact of random measurement errors carried in the coefficient matrix on the precision of coordinate data at the execution end, thus the motion error data was calculated and the coordinate data of the execution end was obtained. With the error data as the sample, the fuzzy neural network model was trained, and the trained fuzzy neural network model was used to predict the error value of Delta-CU parallel robot mechanism. Experimental results showed that the accuracy of fuzzy neural network model prediction was exactly accurate，which would improve the compensation accuracy and applicate in practical Delta-CU parallel mechanism，and provide a practical reference for thermal error compensation of Delta-CU parallel mechanism．Absolute position accuracy was improved from 1.187mm to 0.4mm and the repeat position accuracy was improved from 0.037mm to 0.018mm. The error modeling and analysis method described was reliable and effective, with good compensation effect and obvious accuracy improvement.