Where the Northwest
shapes energy efficiency


Common Considerations in Defining Baselines for Industrial SEM Projects

Created 2/20/2014 by Todd Amundson
Updated 4/1/2014 by BIll Harris
107 views • 2 comments
This is a third whitepaper writen by NW SEM Energy Tracking Savings Protocol team members.  Case studies are presented within to illustrate common issues faced by industrial energy baseline model developers and trade-offs associated with different solutions.  These case studies were written during the latter half of 2013, and finalized in January 2014.
Posted By: Todd Amundson 02/20/14 on 10:24 AM (Pacific Time)
Log in to comment, rate, and share.
Comments (2)
Todd Amundson on 02/20/14 on 10:24 AM (Pacific Time)
The NW SEM ETSP team values your feedback on this whitepaper.  What are baseline energy modeling challenges that you face, and did this paper address any of them?
BIll Harris on 04/01/14 on 10:38 AM (Pacific Time)
Thanks for asking, Todd.  I skimmed the paper, and I have three quick  comments from my experience.

First, I've had some success using mixture models to describe, model, and predict building usage based on 15-minute data.  Distributions of 15-minute power measurements (or, more accurately, short-term integrated power draws) I've seen often look like mixture data--chillers cycle on and off, lighting systems get turned on in the daytime and off and night, ...--and ignoring that makes the analysis harder to conceptualize, at least for me.  That's the biggest point of the three for me.

Mixture models tend to make the estimation of non-production periods rather automatic, too, as they can identify building operating schedules and highlight anomalous building operating states.  Instead of asking for operating schedules as inputs to a regression, I can ask for their intended operating schedules to see if they align with reality as shown by the modeled data.

I've used (mostly) mixtools but also Stan to fit site-meter data.  Mixtools is particularly quick and easy to use.

Second, I'm not convinced that one needs complete samples of cycles of primary and secondary energy drivers including weather, although I do agree that one needs (or at least really wants) samples over the whole range of predictor variables.  Based on my experience so far, a model can be used to develop quite suitable statistical performance with pretty small sets of data, and that's easily testable.  My goal has been to estimate the distribution of time spent in each mixture component and the distribution of power readings in each component.  In some cases, assuming a lognormal power distribution in each component has helped reduce uncertainty.

Then common statistical simulation using the time-series of predictor variables can give the distribution of the adjusted baseline or whatever other measure you want.

Incidentally, ASHRAE 1404-RP is exploring how short a time series can be used to estimate energy consumption, and they seem to be settling in on a number far shorter than a year, the last I checked.

Third, using weekly data, as you note,  may well bypass the issue of autocorrelation with ease at the expense of giving up insights interval data and mixture modeling can provide.  I've experimented with addressing autocorrelation directly using  a hidden Markov modeling approach using depmixS4 or  a state-space approach using MCSim or deSolve and its friends, but I've not used them on real projects yet.

Incidentally,  I'm not stuck just on mixture models.  I've modeled usage at one site where a generalized additive model seemed better suited to detecting and evaluating the contribution of a change to an economizer control system.

Upload a New Version (*.pdf, 30 MB Max)