3.3.1. Shape

The vast majority of approaches take as a continuous regularly sampled window from the time series of the aggregate measurement data input for the DNNs. The range of employed window lengths extends from 90 s [75] to around 9 h [94,98] or even 24 h [66,107]. It is important to note that the number of input samples to the neural networks, i.e., the number of neurons in the first layer, depends on the sampling rate. Extreme values for the size of the input layer are 5 [100], 7 [83], and 10800 [112]. The influence of the window length on the disaggregation performance at a fixed sampling rate is investigated in [52,62,97]. While the investigations have been done on different datasets and sampling rates ([97] → UK-DALE, REDD at 3 and 6 Hz, [62] → REFIT at 10 Hz, and [52] → DRED at 1 Hz), all authors find that the optimal window length depends on the appliance.

Few authors transform the time series data into a two-dimensional representation before feeding it into a DNN. Corresponding publications are marked with 'P2*D*' in the column 'Input' of Table 2: [38,71,123] use the Grammian Angular Field (GAF) [152] to transform a continuous part of the aggregate measurement into a two-dimensional representation which is then fed to a convolutional neural network (CNN). Ref. [38] additionally compares the performance of the GAF with the Markov Transition Field [152] and a Recurrence Plot [153] and finds that the GAF outperforms the other imaging techniques in the vast majority of experiments. Ref. [92], on the other hand, arranged hourly consumption readings into two dimensions by setting the hour of the day as the x-coordinate and the day as the y-coordinate. The authors found that the optimal size of the first filter amounts to 7 × 7, allowing the filter to learn weekly correlations.

A minority of works use a DNN to classify events extracted by a previous detection stage [77,80,114,116] or to classify the on–off status directly from a single time step [82]. These publications are marked with 'na' in the column 'Input' of Table 2.
