Load identification is one of the key technologies in power system planning, operation, and management, playing a crucial role in the efficient scheduling and stable operation of smart grids. Traditional load identification methods typically rely on the closed-set assumption. However, in practical applications, the presence of unknown appliances makes it difficult for algorithms based on this assumption to achieve accurate recognition. To address this issue, an open-set load identification algorithm, OpenAppliance, based on threshold adjustment is proposed. The proposed algorithm integrates deep learning and probabilistic models, calibrating the neural network outputs to enhance the detection capability for unknown categories while maintaining recognition accuracy for known categories. First, load data is transformed into an image format suitable for deep learning, and a CNN-based load identification model is constructed. Then, the OpenAppliance algorithm is applied for post-processing to adjust classification thresholds and optimize recognition results. Finally, the method is validated on the BLUED load dataset and compared with existing load identification algorithms. Experimental results demonstrate that the OpenAppliance algorithm enhances the generalization ability of load identification and significantly improves the accuracy and robustness of the load identification system.