Load profile mining using directed weighted graphs with application towards demand response management

The study of load consumption patterns through subsequence mining is a crucial task in smart cities and buildings as it extracts knowledge to assist in energy planning and management. The existing clustering based solutions do not capture the temporal dependency between the subsequences. Clustering temporally dependent similar subsequences can uncover key insight in a wide range of applications including demand response programs. In this work, we formulate the subsequence clustering on residential building load data as a graph clustering problem on a weighted directed graph. The weighted directed graph structure helps to capture the temporal dependency as well as the similarity between the subsequences. We propose a novel quasi-clique based graph clustering algorithm. No prior information about the number of clusters is required. The comparative study performed on residential building load dataset shows more than 15% improvement over the existing methods. The qualitative analysis of the distinctive patterns obtained from cluster centroids justifies a meaningful cluster set discovery. The labeled subsequences obtained through clustering helps to identify the successively occurring and atypical patterns, measure the stability of consumers that can rank the consumers for their suitability in demand response programs (DR) and load shifting operations.