Classification of energy consumption profiles with a locality-sensitive hashing-based classifier

Abstract

This paper explores the application of a fingerprinting algorithm to the problem of load profile classification in manufacturing environments. Inspired by the Shazam algorithm, which is used for music recognition, we adapt the concept to classify energy consumption patterns in time series data by creating unique fingerprints for each product variant using locality-sensitive hashing (LSH). Our LSH-based classifier (LSHC) demonstrates its potential for efficient and accurate load profile classification, uniquely representing each product variant’s energy consumption pattern. One significant advantage of our method is that indexing just one time series sample per product variant is sufficient for recognizing the same product variant in future queries, similar to one-shot learning in machine learning. Experimental results show that LSHC performs exceptionally well in noise-free or slightly noisy environments, demonstrating robustness to deviations. LSHC is also highly flexible, allowing adjustments in window size, step size, and features based on the data type. However, we identify a potential limitation in the form of index blowup with large training sets, which can significantly increase querying time. While the LSHC meets key requirements such as speed, robustness, and flexibility, challenges with scalability remain. Future work will focus on addressing these scalability issues, extending the algorithm for clustering tasks, and developing methods for real-time identification of load profiles from continuous large time series.