Classification Models and Feature Engineering for Non-Intrusive Load Monitoring

Reducing the utilization of energy is a core strategy for many goals related to environmental sustainability. Understanding which electrical appliances were deployed and how much power was consumed can be a first step toward identifying areas for greater efficiency in electrical usage. Nonintrusive load monitoring (NILM) is a technique for identifying the kinds of electrical appliances that are utilized based only on the measurable power readings. This study develops an improved framework for predictive modeling in NILM applications using the SustDataED2 household appliance data. We propose focusing the analysis only on activated periods of power consumption for each appliance and engineering features that are both predictive and simple to interpret. In the SustDataED2 records, the periods of activation collectively encompass approximately $99 \%$ of the overall power consumption but only account for $18.1 \%$ of the measured records. Within each period of activation, we calculate the following features: duration of activation, peak power consumption, mean power consumption, time to peak power, and the number of local peaks before the maximum value is reached. We also record the first five power readings for the period of activation. Using these features, we construct machine learning classification models in retrospective lookback models and as real-time classifiers. The lookback random forest model achieved a classification accuracy of $97.8 \%$. The real-time random forest models ranged in accuracy from $84.9 \%$ after 1 observation (2 seconds) to $96.8 \%$ accuracy after 5 observations (10 seconds). We further demonstrate that the accuracy of these models is mainly limited by small sample sizes in some classes of appliances. These results improve upon prior research in both their accuracy and their ease of interpretation.