Low accuracy when using DRED

[bontorhumala]

Possibly because of low disaggregation accuracy. Alternative to CO:

• PED (http://dl.acm.org/citation.cfm?id=2887007.2887095)
• Modified CO (https://www.eeh.ee.ethz.ch/uploads/tx_ethpublications/Chrysogelos_SA_2016.pdf)
• Sparse NILM (https://github.com/smakonin/SparseNILM)
• FHMM

[bontorhumala]

22/10: IMPOSSIBLE -> need to detect harmonics, which needs to be sampled at 1.2 - 2 kHz

In modified CO, instead of selecting from all appliances at time t, we only consider set of appliances that have harmonic values > harmonic of aggregated data at time t

Define predicted power of appliances (Model; co_model) and delta (range for each appliance) Define harmonics model

Define A -> set of possible appliance power consumptions Define B -> set of possible harmonics In each loop:

• P = aggregated data
• Extract harmonic of P to I
• Select from A[B > I]

[bontorhumala]

Sparse NILM - see https://github.com/smakonin/SparseNILM/blob/master/piNILM.md

[bontorhumala]

FHMM

• Good Jaccard result (0.7)
• Using features from smart meter and disaggregation, still give bad results (30%). Even worse when disaggregation is normalized (20%). PCA does not help.