To investigate the composition form and evolution law of similar errors between full-scale prototype joints and reduced-scale model joints in steel structure beam-column connections, a full-scale benchmark model test and a 1:2 reduced-scale model test of beam-column end-plate connections were carried out respectively, focusing on the similarities and differences between the hysteretic curves of the two. The test results show that there is an irreparable gap between the test results of the two models without considering the similarity error, and the test results of the scaled model cannot accurately represent the full-scale model. Based on the test results, finite element models are established, and a calculation method to compensate the similarity error of the scale model is proposed. The relative independence between similarity error and test operation error is proved. The similarity error can be accurately simulated by response surface method based on multi-factor correlation. This paper focuses on three test error sources, namely, boundary conditions, plate thickness differences and bolt specification differences, meanwhile the contribution weights of the three error sources and their influence forms on the hysteretic curve are well investigated. Based on the proposed error compensation algorithm, the test results of the scale model are compensated and compared with full scale test results, which proves the effectiveness of the error correction algorithm. The compensated scale model test results are in good agreement with the prototype data, and the error is decreased from 70% to 5%.