Using path signatures to generate individual appliance usage plots from whole-house data
The path signature is a means of feature extraction from multivariate time series data that can be used as input in machine learning models. The objective of this study is to use it in order to be able to tell when an individual appliance is turned on or off, the only data provided being the whole-house monitoring voltage and current graphs, sampled 16kHz. In particular, we generate the usage plot of the fridge on a period of 67 hours. We avail ourselves of machine learning models such as random forest and convolutional neural networks adapted to time series data in order to train our models using past data where each individual appliance is monitored separately. Then, after having trained the model, we make predictions using the unseen whole-house monitoring data - that does not have alongside the individual appliance information - and make them input inside a segmentation algorithm to generate the individual appliance plots. The path signature is doing very well in terms of the accuracy measured as the ratio between the overlapping length of the true versus the predicted plots and the length of the whole interval we make our predictions on.