A deep learning empowered framework for enabling energy savings via non-intrusive load monitoring

This project is in conjunction with an industrial partner named Wifiplug/myma. Their product, a smart plug, can be controlled remotely via an in-house cloud which is interoperable with Amazon Alexa, Apple Siri and Google Home. In their six years of operation, Wifiplug have sold over 30,000 units, and each one transmits its energy readings each minute. Numerous large organisations, including Bosch, Siemens, Procter & and Gamble and Zanussi have expressed interest in the robustness of their product and some are in the process of pilot testing for bespoke use.
This energy data has been archived until now, and the project looks to develop a pipeline of machine learning technologies for this data to monitor the usage, health and likelihood of failure of an electrical appliance from its energy readings and purpose in real time. Presently, little work has been done in the field of power management using machine learning.
We hope to be able to generalise typical appliance behaviour and measure significant deviations as malfunctions/failures. Using laundry machines as an example, an overloaded drum will exhibit higher readings in the spin and rinses, which we will be able to detect. This begins with a categorical classifier using a deep recurrent neural network to determine the type of appliance when it is switched on, trained on the existing dataset. There are some complications with the energy readings; most notably a heavy imbalance in types of appliance. For example, there are significantly less dishwashers connected to smart plugs than laundry machines. We must assess and apply data science/analysis techniques to compensate for this in model training. Furthermore, as part of the original brief, usage habits of the appliance can be monitored on edge. Using predictive analytics, we hope to be able to determine when an appliance will enter a prolonged state of idleness and power down the appliance to conserve energy.

The industrial partner is currently developing an enhanced version of the smart plug, capable of capturing 50 energy readings per second as opposed to 1 per minute. This should offer greater precision in behaviour analysis than the existing archives, but we must investigate whether and how the archived readings can be best utilised in the machine learning pipeline for the new hardware.
To summarise, Wifiplug have vast archives of energy data and the means to continue producing it in far richer detail, yet currently have no way of leveraging it for their customers. This project looks to develop a comprehensive, intelligent pipeline to gain a new depth of understanding in appliances, from failure detection to energy conservation.