To control upper-limb exoskeletons and prostheses, surface electromyogram (sEMG) is widely used for estimation of joint angles. However, the variations in the load carried by the user can substantially change the recorded sEMG and consequently degrade the accuracy of joint angle estimation. In this paper, we aim to deal with this problem by training classification models using a pool of sEMG data recorded from all different loads. The classification models are trained as either subject-specific or subject-independent, and their results are compared with the performance of classification models that have information about the carried load. To evaluate the proposed system, the sEMG signals are recorded during elbow flexion and extension from three participants at four different loads (i.e. 1, 2, 4 and 6 Kg) and six different angles (i.e. 0, 30, 60, 90, 120, 150 degrees). The results show while the loads were assumed unknown and the applied training data was relatively small, the proposed joint angle estimation model performed significantly above the chance level in both the subject-specific and subject-independent models. However, transferring from known to unknown load in the subject-specific classifiers leads to 20% to 32% loss in the average accuracy.