As a result of a number of recent events around the world, energy prices are on the rise in a big way. In the United States, the U.S. Energy Information Administration reported that the average price for residential electricity rose by 8% in 2022. That is a bitter pill for consumers to swallow, but on the world stage, it is a relative drop in the bucket. In much of Europe, for example, prices for oil and natural gas, which are often used to generate electricity, have risen by as much as 60% and 400%, respectively. For many, price increases of this magnitude are more than just an inconvenience and may cause serious problems as winter sets in.
One area of opportunity that may help to alleviate some of the pain we are experiencing is in reducing the amount of energy that we consume. But we just use the energy that we need, so how do we do that? Well, there may be hidden, pain-free opportunities to reduce energy consumption that we are just not aware of. Non-Intrusive Appliance Load Monitoring (NIALM) is a technique that can continuously monitor which appliances are running, and when, in a residence or small office building. Information from NIALM systems can provide insights about appliances that may be running, and consuming substantial amounts of energy, when they are not in use and could be switched off.
There is a problem with existing NIALM implementations, however — they require many hundreds of labeled power signal readings for each and every appliance that you want to recognize, and from each of their operational states (for example, a washing machine during fill, spin, and drain cycles). Collecting and labeling such an extensive dataset is labor intensive and costly. This data is then used to train a machine learning algorithm, and the more data that is included in the training process, the more expensive it is in terms of computational time. Unfortunately, these issues are factors preventing the widespread adoption of this technology.
There may be a better path forward, however, according to a team of researchers at the University of Johannesburg. They have developed a novel few-shot classification pipeline that can be trained with less than 10 labeled samples per class, and yet is capable of achieving a high degree of accuracy in NIALM tasks. Their primary algorithm borrows from both Siamese networks and prototypical few-shot networks. Siamese networks excel at producing one-shot and few-shot networks, however, appliance activation periods vary greatly, which leads to an imbalance in the number of samples collected for each class. This imbalance between classes can lead to inaccuracies in model predictions.
To address issues with data imbalance, the team also borrowed from prototypical few-shot networks. Since they only produce an average value embedding point of the examples in each class during training, problems arising from certain classes containing more samples than others disappear. To further speed up the training process, the team performs what they call a similarity test before training. Running this on the entire training set helps to understand the level of quality of each sample, and discard any informative or misleading data before training the model.
To test their methods, the researchers created a NIALM dataset for fourteen types of appliances (e.g.: desktop computer, refrigerator, two-plate stove), then trained their network using samples from 10 of those appliances. A total of 10 samples were included in each class. The model was then tested on four appliances it had not previously been trained on. Given 10 samples, the classification accuracy climbed to as high as 97.83%. Even with just a single appliance sample, the classification accuracy achieved was greater than 91%. These tests proved that the team's method was quite capable of performing NIALM recognition tasks with a high level of accuracy.
Moving forward, the researchers plan to use their methods to develop practical systems that can detect high-value appliances that are not working properly by using a single monitor in a building. This has applications in predicting equipment failures before they happen. By identifying failing equipment early, maintenance can be performed that saves both time and money down the road.