Load identification technology can effectively distinguish different electrical types. This is of significance for power consumption strategy formulation and demand response. There is a problem in that the present load identification technology cannot effectively realize the load feature fusion and the combination of different recognizer models. Thus an integrated learning method of residential power load identification feature image based on heterogeneous basis Stacking mechanism is proposed. This method realizes feature fusion by constructing feature images and fully excavates the deep-seated features contained in feature images using a convolutional neural network, so as to solve the problem that the traditional methods are not deep enough in feature mining. At the same time, the ensemble learning Stacking method is introduced to combine a variety of heterogeneous load identification models to integrate the advantages of various models and overcome the problem of the simplification of traditional methods. Finally, the public data set PLAID is used to verify and complete the engineering application on laboratory electrical equipment. The results show that this method has high recognition accuracy and application value. This work is supported by the Joint Project of the Commission of National Natural Science Foundation of China- Smart Grid of State Grid Corporation of China (No. U2066212).