Message from the editor

Cristina Rottondi, Dalle Molle Institute for Artificial Intelligence 

Message from the Chair

The world is changing again. Computer architecture faces the end of Denard Scaling, moving the burden of energy efficiency toward software developers. Smart cities and smart grids are real and increasingly common, but they struggle to efficiently corral and analyze data into actionable insights. Sustainability is bigger than solar or wind energy.  Moving forward, sustainable computing solutions must consider long-term effects on society.  Are our solutions ethical?  Are human, ecological and energy resources sustainable for long-term autonomous deployments?  Have we discovered the best, sustainable solutions?  We haven't addressed these problems yet. It is fair to wonder aloud. Is our STC ready to address these challenges?

It is time to restructure our STC to better support leaders at the forefront of the new challenges in sustainable computing.  We will recruit a new board, senior advisors, newsletter contributors and research pioneers.  We are looking for fresh faces and established members.  This STC must continue to lead by highlighting, promoting and fostering the best research and engineering in sustainable computing. 

Over the next few months, as we restructure, the newsletter will go dark. In 2019, our STC will reemerge ready to take on new challenges facing sustainable computing.  I will reach out to the full membership to plan workshops and request feedback on our restructuring plans. If you would like to be more involved, please email me. We need as many volunteers as possible. I am truly excited about the future for sustainable computing and the important role this STC will continue to play.

Thank you,

Dr. Christopher Stewart

Occupancy Inference based on Machine Learning Techniques

The knowledge of occupancy in domestic environments is vital for many applications, such as energy management, building management, demand/response, security, etc. But, in the majority of the cases, it is very difficult or even impossible to install sensors for receiving occupancy information.

Thus, the occupancy (absence – presence or two-class classification scenario) should be detected and inferred utilizing machine learning techniques via electricity and water consumption data received from smart meters in a domestic environment.

To this end, the power consumption of crucial electrical appliances (i.e. TV, washing machine, refrigerator, and hair dryer) is monitored every minute. Also, four water consumption sensors have been installed for monitoring the water usage by the occupants in the kitchen as well as the water consumed by the dishwasher and the washing machine. Finally, an occupancy sensor has been installed at the main entrance of the building detecting entries and exits and measuring its occupancy. This information is utilized as ground truth.

After retrieving the raw data of the three systems, a processing step was performed in order to create the final dataset which includes events per 1-minute intervals of all the measured features. The initial aggregated dataset constructed after processing the raw data contains 9 features [Central Power (lights of the domestic environment), Refrigerator, TV, Washing Machine, Dryer, Cold Water - Kitchen, Hot Water - Kitchen, Dishwasher - Water, Washing Machine - Water] denoted hereafter as [CP, R, TV, WM, D, CWK, HWK, DW, WMW] and the target Occupancy, denoted hereafter as [OCCUP]. The dataset contains energy and water consumption data of 1-minute resolution for a time interval of 16 consecutive days during summer time. Thus, the shape of overall dataset is 23040x9 (without taking into account the target feature) and its sparsity is 74.44%.

In order to rank the influence of each feature to occupancy inference and extract the more useful information, we have used Mutual Information (MI) as the feature selection technique. MI measures how much one random variable provides information. It is a dimensionless quantity, and can be thought of as the reduction in uncertainty about one random variable given knowledge of another. Thus, we decide to use only the top-5 ranked features for occupancy inference, meaning Central Power, Cold Water – Kitchen, Washing Machine, Refrigerator and Washing Machine – Water. Under this condition, the shape of overall dataset is 23040x5 (again without taking into account the target feature) and its sparsity is reduced to 70.76% (from 74.44%).

In the sequence, machine learning techniques have been utilized in order to infer the occupancy in a building. The tested machine learning algorithms are:

• Support Vector Machines (SVMs);
• Decision Trees (DT);
• Random Forest (RF);
• Back-Propagation Network (BPN).

All these approaches were also combined with the AdaBoost algorithm for even more accurate performance.

Our main objective is to find the predictive model that is more efficient on occupancy inference based on energy and water consumption data. To that end, our simulation schema is based on the application of all tested classifiers and their boosting versions on both Initial-DS and MI-DS.

The overall framework has been thoroughly evaluated in [1].

References: 

1. T. Vafeiadis, S. Zikos, G. Stavropoulos, D. Ioannidis, S. Krinidis, D. Tzovaras and K. Moustakas, “Machine Learning Based Occupancy Detection Via The Use of Smart Meters”, International Conference on Energy Science and Electrical Engineering (ICESEE 2017), Budapest, Hungary, 20-22 October, 2017

Dr. Sabidur Raman, Dr. Abishek Gupta, Prof. Biswanath Mukherjee, University of California Davis, USA

Prof. Massimo Tornatore, Politecnico di Milano, Italy