Correlation between static features and HVAC

[nipunbatra]

Why: We observed that the HVAC performance for Tensor factorisation becomes worse (CASE 2) when we add in the static features. Why is it so?

How: We will look into the correlation between static factors and the HVAC energy consumption. If this correlation is close to 0, the learnt latent factors should be 0 to ensure these features don't contribute. Is this the case?

[yilingjia]

Is this what you mean? The correlation between static factors and the HVAC energy consumption.

[nipunbatra]

This is the first step. Next:

1. Do a scatter between HVAC energy and static features. This is to see if a few outliers are causing low correlation
2. Pass each of these static factors to learn_HAT one at a time. Keep a=2. This way, we have one fixed latent factor for H and one will be learnt. Since correlation with static factors is low, we would hope to see that the corresponding latent factor for appliance would be close to 0.

[yilingjia]

Do a scatter between HVAC energy and static features. This is to see if a few outliers are causing low correlation

This is the result I got by scattering HVAC energy and static features. For this,

1. the energy consumption is related with the area, but seems they have different gradients, somehow the shape like a sector.
2. For the number of rooms and the total occupants, the lowest energy consumption seems related with them, but comparing with the area, the points are more scattered. One thing about this is for the number of rooms, seems (hvac_5, hvac_6, hvac_7)have similar pattern while the other three have another pattern.
3. From the scatter figure, we do observe some outliers, we can do some experiments to check the correlation between them without the outliers.

[yilingjia]

Pass each of these static factors to learn_HAT one at a time. Keep a=2. This way, we have one fixed latent factor for H and one will be learned. Since correlation with static factors is low, we would hope to see that the corresponding latent factor for appliance would be close to 0.

About the second one, I am not sure if I understand it correctly. I just draw a figure for the tensor decomposition blew.

I am not sure about the "corresponding latent factor" you mentioned. Based on the decomposition, I think we need to consider the factors learned in tensor A and matrix T to check if the static factors was related with the consumption or not. Is my understanding correct?

Thanks!

[nipunbatra]

Yes. That's right.

[yilingjia]

I did experiments you mentioned before and got the results:

1. The energy consumptions are related with the area, but not so related with the total occupants or the number of the rooms. The heat maps show the latent factors for this observation.
2. The static factors are more related with the aggregate consumption than hvac consumption.