Smart Grids

Power grids are transitioning to smart grids, partly to reduce CO2 footprint.

This includes deployment of renewable energy sources (RES), e.g., solar and wind energy. RES incur major challenges in ensuring that load is always balanced by production. Demand response (DR) algorithms are a key (part of the) solution to this balancing problem, aiming to steer power consumption. To successfully deploy DR, it is of prime importance to thoroughly characterize (and possibly predict) consumption patterns, and especially what part can be shifted in time (e.g., to match RES production).

This characterization requires advanced data analytics, and machine learning algorithms are highly valuable both to automatically detect/learn power consumption behavior as well as control it (e.g., in RES balancing).

Some of the topics we recently have been working on in this area include the following:

  • Scalable load profile clustering: in contrast to state-of-the-art that a priori imposes different classes (e.g., week- vs weekend days), we study unsupervised algorithms that do not make any a priori assumptions about such classes, and adopt feature representations to improve scalability of the algorithms.
  • Quantitative flexibility assessment based on real-world data: we characterize behavior of users with statistical models, and derive the flexibility (i.e., over what time frames we can shift power consumption, and what load volume that amounts to) that is exhibited in such observed electrical load patterns.
  • Non-intrusive load monitoring (NILM): we are working on improved NILM algorithms, i.e., decomposing total consumption measurement into individual device consumption. We investigate new features (e.g., inactive current, binarized VI trajectories), and explore the potential of recent advances in machine learning for device classification (e.g., using deep learning).

Complementary to this data analytics and machine learning line of work, we also contribute to communication software approaches, e.g., based on information centric networking (ICN) approaches such as in the C-DAX project.

Staff

Chris Develder, Tom Dhaene, Matthias Strobbe, Dirk Deschrijver

Researchers

Nasrin Sadeghianpourhamami, Leen De Baets, Joeri Ruyssinck.

Projects

Key publications

Two-stage clustering of power consumption profiles: (i) within single households, we first identify categories of “typical day” profiles, (ii) we then cluster together users that share similar “typical day” profile mixtures.
Two-stage clustering of power consumption profiles: (i) within single households, we first identify categories of “typical day” profiles, (ii) we then cluster together users that share similar “typical day” profile mixtures.

Analysis of EV charging flexibility exploitation in two demand response (DR) scenarios: load flattening and renewable energy source (RES) balancing. The top figure (a) shows power consumption profiles in the uncontrolled business as usual (BAU) scenario vs the two DR cases. The bottom figures show the amount of BAU power that is shifted over what amount of time in case of (b) load flattening and (c) RES balancing.
Analysis of EV charging flexibility exploitation in two demand response (DR) scenarios: load flattening and renewable energy source (RES) balancing. The top figure (a) shows power consumption profiles in the uncontrolled business as usual (BAU) scenario vs the two DR cases. The bottom figures show the amount of BAU power that is shifted over what amount of time in case of (b) load flattening and (c) RES balancing.