Non Intrusive Load Monitoring: Systems, Metrics and Use Cases

Buildings across the world contribute significantly to the overall energy consumption. Targeted feedback can help occupants optimise energy consumption. In our first work we present techniques for actionable feedback across fridges and air conditioning (HVAC) units, which can save upto 25% of fridge energy and identify homes needing feedback on HVAC setpoint schedule with 84% accuracy. In our next work, we do an extensive sensor deployment in a home in Delhi, India; monitoring appliance level power, home aggregate power and other ambient parameters. Our study presents various insights unseen in the developed world, such as: frequent voltage brownouts, poor network reliability, long lasting blackouts, heavy dominance of fridge and HVAC to overall energy consumption. Our study verifies that measuring appliance level power scales poorly in cost and maintenance. Non-intrusive load monitoring (NILM) is viewed as a viable alternative where machine learning techniques are used to break down aggregate household energy consumption into contributing appliances. Despite the existence of a rich volume of literature in NILM, it remained virtually impossible to compare NILM works due to: i) lack of existence of benchmarks; ii) previous work tested on single data set; iii) inconsistent metrics. To address these challenges we developed an open source toolkit: Non-intrusive load monitoring toolkit (NILMTK), designed specifically to enable the comparison of NILM algorithms. While many new NILM techniques have been proposed in recent times, it is not clear if these can enable energy saving and whether higher accuracy translates to higher energy saving. We explore these questions in our recent work and find that existing energy disaggregation techniques do not provide power traces with sufficient fidelity to support the feedback techniques we developed in our earlier work. Our results indicate a need to revisit the metrics by which disaggregation is evaluated.