Thinking about HW and also looking at the Excel trendlines using moving average of last 20 thru a few hundred points I suspect we should take 95% of the weights from the last few hours rather than the last 10 hours since there are quite large changes (20-30% for Caltech from 9:00am to 12:00noon) in say a 3 hour interval. Three hours would allow some consistency with the plateau parameters and also should not cause a large loss of synchronization due to the averaging (you should probably check this). Also looking at the smoothing effects of a moving average over say the last 3 hours (say 60 points at 3 mins interval between observations) that looks like a reasonable interval to get good smoothing.
The seasonal effect was chosen to match the weekly cycle, i.e. seven days.