Load decomposition is an important means to obtain electrical details of appliances and analyze users’ power consuming behavior, which helps to strengthen demand side management of smart grid. Since current non-intrusive load decomposition researches lack concern about power consumption patterns of appliances and weak migration ability of these models, a decomposition method for appliance load based on power consumption patterns and dictionary learning is proposed. Typical consumption patterns of appliance load are extracted by clustering. According to the consumption patterns contained in the appliances from testing house to be measured, dictionay learning algorithm is used to train pattern dictionary of the appliances. Then sparse representation is applied to total load with pattern dictionary to realize load decomposition. Accuracy of the proposed method and its performance on house migration are verified by decomposition results on test dataset.