On the Impact of the Sequence Length on Sequence-to-Sequence and Sequence-to-Point Learning for NILM

The Sequence-to-Sequence (S2S) and Sequence-to-Point (S2P) optimization methods achieve remarkable accuracy results for load disaggregation tasks. Internally, they rely on neural networks, trained to identify the power consumption of a single appliance under consideration from a sequence of aggregate power data. Their most important configuration parameter - the number of input data samples to consider - is, however, mostly set to a fixed value. As a result thereof, the amount of historical data available at the algorithm's input is governed by the sampling interval of the used input data. For example, UK-DALE [5] provides samples every 6 s, so a sequence length of 599 samples (as proposed in [9]) makes approximately 1 h of historical data available to the disaggregation algorithm. No analyses of the impact of the sequence length on the NILM performance have been documented in literature to date. We hence present a methodological assessment of the sensitivity of S2S and S2P to variations of their input sequence length parameter. Our results show that setting a per-device parameter value leads to improved disaggregation results; however, the required values need to be determined empirically, as they are unrelated to the appliances' operational durations. Even if only a single value may be set, an informed choice (rather than using the default value) can drastically improve NILM performance.