In the context of the current smart grid, there are numerous typical low power factor load scenarios. The feature differences among different scenarios are small, and the structure of power load data is complex. As a result, classifying low-power electrical loads has always been a difficult problem in practical research. It is necessary to develop advanced models to improve the accuracy and efficiency of classification. This paper combines cluster analysis and classifier recognition, and attempts to conduct analysis and implementation from the combination of power load curve cluster analysis based on the t-SNE algorithm and improved K-means, and load pattern recognition based on the support vector machine classifier. The t-SNE algorithm can not only reflect the local sensitivity of the original data but also retain its global structural features, and can be effectively applied to load data with a low power factor. The improved K-means uses the elbow criterion to determine the number of clusters K. Selects the initial center points using a method based on the density and dissimilarity attributes of the data set, which can effectively improve the computational efficiency, accuracy, and cluster stability. The SVM classifier can fully utilize the clustering results and features. Once the classifier is trained, it can quickly perform intelligent classification and recognition on new unknown load data, thus improving efficiency. This paper evaluates the effectiveness and stability of the clustering effect of the model from validity indicators such as SC, CHI, and DBI, and all obtain good results. Moreover, the classification accuracy of the SVM classifier on the test set reaches 100%.