A Kaizen Like Algorithm for Iterative PMV

One of the interesting conundrums about the PMV calculation is that it's a predicted mean vote. And it was (presumably) built around statistics gathered decades ago. However, the PMV itself is really a subjective criteria. And subjective criteria are often best quantified and correlated with real world data.

An interesting and powerful strategic advantage we'll have is that we're gathering time series data about indoor environmental conditions and can directly correlate actual subjective ratings from users with measured indoor conditions (rather than infer them algorithmically from outdated experimental models).

So imagine if we have a reading:

Sensor Reading n:
Timestamp Tr = 1300816141 = GMT: Tue, 22 Mar 2011 17:49:01 GMT
Temp C = 24
Relative Humidity RH = 40%
CO2 = 400ppm
CO = 2ppm

Then, we have an *optional* application we provide each client that allows them to have the building occupants periodically ask them what their comfort level is. It's a simple, non-intrusive application; something that justs asks to them to rate their comfort level. We can use the traditional Fanger 7 point index or make our own, it really doesn't matter. We can even ask them simple PMV related questions such as if they're heavily dressed, etc (although these are somewhat irrelevant). In any case what we end up with is a quantity of actual comfort votes we can use to correlate with measured indoor conditions at that time interval. A primitive pseudo-algorithm might look something like this:

Comfort Votes [1,2,3,4,5...n]:
Timestamp Tc_1 = 1300816141 = GMT: Tue, 22 Mar 2011 17:49:01 GMT
Comfort Rating C_1 = +2
Timestamp Tc_2 = 1300816141 = GMT: Tue, 22 Mar 2011 17:49:01 GMT
Comfort Rating C_2 = +3
Timestamp Tc_n = 1300816141 = GMT: Tue, 22 Mar 2011 17:49:01 GMT
Comfort Rating C_n = +3


Then we can calculate a new mean:

for each Comfort Vote @ Timestamp x {
Sum = Sum + C[i]
Mean = Sum / n

Then we have a concrete Mean for Sensor Reading n @ Timestamp x. As we gather more votes over long periods of time, we can regress over the historical Actual Mean Votes (AMV) and perform an analysis of variance in the measured variables at those instances in time. So, if we see the AMV changing as a function of CO2 for instance (with little variance in other values) we can create new (and more accurate) phi functions for CO2 changes. Thusly, we can have a continuously adapting PMV algorithm that more accurately determines occupant comfort. We can also extend and adapt the algorithm to various sub populations (school kids, office workers, etc) since many sub populations perceive thermal comfort differently.