Big difference between the actual power data & dis-aggregated power data

[Abdul-Samad-Saleem]

I have plotted the sum of the dis-aggregated 15 loads obtained in gsp_result data frame. The actual aggregated data is also plotted to compare the results with the actual data. There appears to be a big difference between the actual aggregate power and the result obtained using your code. Can you verify?

[loneharoon]

This is possible. Please tune the parameters: sigma, ri, T_Positive, T_Negative, and instancelimit, after reading their meanings from the paper. Also, if you go into gsp.feature_matching_module() function, you will notice that it drops few appliances for which the stop edge is not found. The stop edge gets often missing due to lower sampling rate. Also, I remember there is a parameter which defines the number of clusters. Changing that affects the results too.

[Abdul-Samad-Saleem]

I will try testing with different tuning parameters. Regarding other comments relating to missing appliances: I understand that the dis-aggregated energy is already much higher than the total power input to the algorithm. If the missing devices were included the error would increase even further. It appears to me that the rising edges of high power appliances could possibly be paired with falling edged of some different appliances causing false stretches in their operation time hence increasing the dis-aggregated power.

[loneharoon]

That is possible. I have verified the results of my Python version with the author's Matlab version at the time of it's implementation. And my output was exactly the same as of his. Please contact the paper author's for the issue.