After looking at yesterday’s data, I think I agree with NEA that you need at least a 3-hour average reading to make it reliable. It’s the minimum samples that you require to have a meaningful data. The 1-hour estimates simply fluctuate too much (or the model is too simple, perhaps it can be modelled better from the 3-hour reading). The ‘lag’ problem still persists, though, so I suppose there is still room for one more variable to indicate the ‘current trend’ of the data and have some predictive value to it.
Update: Apparently the problem lies in the piecewise linear relationship between PM10 concentration and PSI value, as highlighted in this excellent post.