I attended a seminar at school this afternoon, given by Dr. Matteo Giuliani from Politecnico di Milano. To me, the SmartH₂O project is interesting in several aspects:
The use of disaggregation algorithm: a smart meter was installed at each household in the study. The meter is able to log water consumption in fractions of a second, but what it logs is the total consumption. The study, however, seeks to breakdown this consumption to different usages such as showering, toilet, laundry, and so on. To achieve this, the researchers developed a disaggregation algorithm. My question was where did they get the data to train such an algorithm? As it turns out, there are two possible approaches.
- The first one is to install a meter on every equipment in the house. This approach is intrusive, and they could not do it for this project, but they managed to find a dataset in the the literature. Someone in Canada was doing something similar for his PhD, so installed meters in his own house and collected data for an entire year. The SmartH₂O team was able to test their algorithm against this dataset.
- The second approach is to ask users to keep a diary and note down every time they run a single appliance. Certainly, this approach is less accurate. In addition, the diaries are submitted only days or weeks later, by which time it is too late to correct any inaccuracies. To deal with this issue, the SmartH₂O team is developing an app so that the participants can log their usage in real time and the system can cross check these entries as soon as they are keyed in.
I should keep this in mind in case I’m doing something similar in the future.
User segmentation: the researchers used unsupervised learning to classify their users. They found out that the most simple and useful classification is to use two indices:
- Consumption level: very high, high, medium, and low.
- Peak: weekday or weekend.
Using this system, they were able to divide the consumers into 8 different groups.
Gamification and behavioural economics: To increase consumer engagement, the project team created an online platform so that users can track their consumption, earn points for water savings and compare themselves with their neighbourhood. They even launched a board game so that children can participate too. Points earned can be exchanged for gifts such as water saving features. This is another demonstration that feedback and consequences work better in modifying behaviour than antecedents.
The SmartH₂O project is interesting in both technical (in particular, machine learning) and socioeconomic aspects. As it turns out, PUB is doing the same thing in Singapore. I look forward to seeing the results. It would be best if they engage SUTD too.