Improving K-means++ algorithm with GMM based on shapelet temporal features and combined weights

The rationality and accuracy of selecting grid operation modes are crucial for ensuring the safe and stable operation of power systems. With the increasing complexity of grid structure and the intermittency and uncertainty of new energy and flexible loads, the operation scenarios of the power system show significant temporal features. The current traditional scene clustering methods do not adequately consider temporal features, and the manual extraction of typical operation modes is highly subjective, which bring new challenges to the safe and stable operation of the power system. Therefore, this paper proposes an improved K-means++ algorithm for Gaussian mixture model (GMM) based on shapelet temporal features and combined weights. Firstly, the Shapelet algorithm is applied to extract the typical features of multidimensional time series variables. Secondly, the UMAP algorithm maps the complex high-dimensional features to a lower-dimensional space, achieving feature dimensionality reduction. And in order to fully consider the degree of difference and correlation of temporal features, the improved entropy weight method fused with the CRITIC weight method is used to assign the combined weights to temporal features. Finally, the weighted Gaussian mixture model (GMM) is used to improve the K-means++ algorithm for clustering the grid operation data of a province. The results of the algorithm show that compared with the traditional scene clustering method, the method in this paper can effectively extract complex temporal features, and has better performance in silhouette coefficient, DBI and CHI indexes. Moreover, the derived typical operation modes for the power grid prove to be highly illustrative and representative.