Next I usually compare my Sense hourly results against my PG&E results. Not surprisingly Sense results are usually very reliably within 2% of my revenue utility meter but I still the check regularly for continued confidence building. There are 3 steps to this once I have double-checked that all the data is there (Task 1):
- Rolling up my PG&E CSV data which comes in 15 minute increments into hourly increments.
- Converting everything to net usage measurements since I have solar and that is the only basis
for comparison. PG&E can’t see my solar production, so all the meter provides is net readings, Import and Export. Here’s a view of the hourized PG&E data, also converted to Net (Import - Export). Notice that I also keep around a field called Type to tell when the information from PG&E is measured or estimated (when there are connection problems from the meter to the mothership, the mothership estimates)
- Joining the data by DateTime hour, then comparing via a graph. The 45 degree unity graph shows that there is almost an exact linear correlation between my utility measurement and Sense, except one hour where Sense came up a little short, circled in red.
I have automated this 3 step task, in an R script that takes about 5 seconds to run. But it can be done in about 20 minutes by an experienced Excel user.
I typically do one more piece of analysis on the data, to better understand the linear relationship. It looks like a 45 degree line running through zero, but if I do a linear regression fitting, I find that both those assumptions are off by a little bit. I’ll just highlight three numbers I the linear fitting below.
The bottom line is that
Sense Net = (1.0091171 x PGE Net) - 0.0010101 kW
That means that the Sense combined net measurement (Total Usage - Solar Production) is generally almost 1% larger than the PG&E combined net measurement (Import - Export) though also offset by -1W. The R-squared of 1 highlights that this is almost a perfect linear fit, even with the one slightly off-linear point.
BTW - here’s what caused that one non-linear point - a dropout around 4AM on April 27th.
Now let’s see how things works using Perplexity Pro AI. I’ll start a new post for that.
