The quest for automatically tracking electrical appliance use
At Western Power, we are beginning to apply machine learning techniques in a bid to improve efficiency. So far, there have been several successful test cases:
Identifying conductor defects using the DataRobot platform
Detecting solar PV installations from aerial photographs using Convolutional Neural Networks
Predicting power pole top fires using decision trees and Xgboost
In this use case, we are attempting to identify electrical appliances being utilised behind customer meters from high frequency interval meters. That is, electricity meters that record the amount of kilowatt hours of electricity consumed at daily, hourly and five minute intervals.
A few reasons for attempting this are:
We can help manage critical peak demand periods by identifying areas of demand that exceed local network constraints.
Help customers understand which appliances are contributing the most to their utility bill. This information is useful for deciding when to replace under-performing appliances and cut down on needless waste.
Develop a better understanding of electricity consumption patterns, leading to improved demand forecasting. This is useful for long-term network capacity planning.
Applying machine learning methods to this task means we can reduce the cost of monitoring and upgrading Western Power's electricity network.
The first challenge with any machine learning project is to capture the data. You might think this would be easy for an electricity utility to achieve. However, there are several important constraints that make this difficult:
Most electricity meters are only manually read once every two months when a meter reader visits each home or business and reads off the accumulated meter reading. This means we typically only obtain a few readings a year.
Where we do have interval meters, there are strong privacy laws that make accessing and sharing the data difficult.
Fortunately, data loggers are becoming increasingly common and relatively inexpensive. We installed a data logger from Wattwatchers (see https://www.wattwatchers.com.au/) called the Auditor 3. This is a Wi-Fi enabled meter that can be installed by a licensed electrician in a typical meter box.
The diagram above illustrates how this works. In a nutshell, we get data uploaded to the cloud from participating electricity consumers.
If you would like to participate in the challenge, the resulting data can be inspected here: https://coble-neal.shinyapps.io/App-4/
There is a wealth of information that can be readily gleaned from the data.
For example, the chart above shows that at a daily interval, there is a clear seasonal cycle with frequent spikes. Overlaying maximum daily temperature recordings indicates a correlation, suggesting that many of the spikes in electricity consumption are likely caused by extreme temperature.
From this, it would be reasonable to assume that there is an air-conditioning system in use. The next chart shows the same data at an hourly interval. The grey vertical bands correspond to weekends. This chart suggests a weekly cycle.
Drilling down to inspect a single day reveals a daily cycle, as indicated in the chart below.
Finally, using five minute interval data reveals distinct electrical appliance profiles.
At this point, we can that there are several appliances in use at different times of the day. The challenging aspect of the data is that the use of multiple appliances simultaneously changes the profile. Analyts can readily distinguish between these profiles when inspecting charts like those presented here. But how can a this task be assigned to a computer to automatically detect and catalogue appliances?
Choice of algorithms to solve this problem
Successful automation of appliance use detection requires restructuring the data so that it can be presented in a consistent, well defined manner. And there is a variety of alternative methods to achieving this.
In our first attempt several years ago, we attempted a rules-based approach. This is implemented by establishing a series of hand-crafted rules in a logical sequence as illustrated in the flow chart below.
This approach proved successful with a relatively high accuracy rate. One snag, however, is that Western Power has over a million meters to monitor. Hand crafting solutions for each meter is too time consuming and costly.
In our second attempt, we enlisted the help of the Perth Machine Learning Group. This time a combination of algorithms were employed to develop a viable solution.
Sherief Khorsid at Three Springs Capital suggested an unsupervised learning method to identify clusters. The algorithms sorted spikes by size and frequency as indicated below.
This provided the ability to rebuild the original aggregate profile data by appliance with each appliance profile identified as a specific class.
As you can see, this approach was largely successful in identifying distinct appliances operating.
This foray in machine learning highlighted a few lessons and ideas for further research:
Use supervised learning Instead of k-means. IE
Use a classification Neural Network or XGBoost algorithm to classify the each of the peak phases. This will involve manually labelling the peak phases.
If you would like to get involved in the challenge please download the data and join the automatically tracking electrical appliance challenge in the forum
Grant Coble-Neal is Insights & Analytics Manager at Western Power where he is driving the development of the next generation of advanced analytics. Leading a core team of 10, Grant is passionate about finding ways to value-add. A career economist specialising in investment appraisal, Grant has assisted Western Power with key investment evaluations.
He is also championing the ethical development of machine learning for the benefit of the community that Western Power serves.
More information about Grant is available on his LinkedIn profile https://www.linkedin.com/in/gcobleneal/