Post-processing for event-based non-intrusive load monitoring

Most current non-intrusive load monitoring (NILM) algorithms disaggregate one appliance at a time, remove the appliance contribution towards the total load, and then move on to the next appliance. On one hand, this is effective since it avoids multi-class classification, and analytical models for each appli- ance can be developed independently of other appliances, and thus potentially transferred to unseen houses that have different sets of appliances. On the other hand, however, these methods can significantly under/over estimate the total consumption since they do not minimise the difference between the measured aggregate readings and the sum of estimated individual loads. By considering this difference, we propose a post-processing approach for improving the accuracy of event-based NILM. We pose an optimisation problem to refine the original disaggregation result and propose a heuristic to solve a (combinatorial) boolean quadratic problem through relaxing zero-one constraint sets to compact zero-one intervals. We propose a method to set the regularization term, based on the appliance working power. We demonstrate high performance of the proposed post-processing method compared with the simulated annealing method and original disaggregation results, for three houses in the REFIT dataset using two state-of-the-art event-based NILM methods.