An embedded deep-clustering-based load profiling framework

Abstract

• A novel load profiling framework using deep clustering to for feature reduction. • Framework eliminates preprocessing in hybrid load profiling algorithms. • Autoencoders capture main non-linear features to separate different load groups. • Improved SI on real-world big data compared to classical clustering techniques. • Framework produces robust load profiles that describe distinct customer groups. Load profiling is an essential step in several smart meter analytics tasks, such as forecasting and planning, that directly impact sustainable energy management. As most real-world smart meter data are unlabeled, the unsupervised learning clustering-driven load profiling approach aims to group related customers based on usage trends. However, due to the curse of dimensionality, conventional clustering algorithms perform poorly and often lead to inadequate load curves. The proposed load profiling framework’s novelty is twofold. First, using an autoencoder the framework represents data as a hierarchy within the layers of the deep network, allowing for dimensionality reduction and highly nonlinear decision separation between clusters at the autoencoder bottleneck level. This is achieved by implementing autoencoder-based clustering that automatically converts smart meter data into more clustering-friendly representations that retain the original data characteristics. Second, the framework integrates dimensionality reduction and clustering into a single end-to-end unsupervised learning framework. The deep clustering framework is then tested on two real-world smart meter datasets, and the experimental results show that the proposed framework produces significantly better load profiles when compared to classical clustering algorithms as well as previous hybrid frameworks proposed in the literature. The paper concludes with a discussion of the results and a set of future directions for improving the load profiling task.