The electric power (in Watt) as reported by S0 meters is very inaccurate.
I noticed that when the actual consumption is very low, in my case below 6W using a 2000 counts/kWh counter, the reported power is far off the actual power and often as high as 180W. With 1000 counts/kWh this can go up to 360W.
I looked at the source code - S0MeterBase::ParseLine() - and I suspect the problem is due to 2 effects:

1. at low power there may be fewer than 1 pulse per 30 intervals (5 minutes), the idea is to report this as 0W. It also resets the interval counter to 0. Now when a new pulse arrives (which would be the 31st interval) it is considered to be in the first interval and the actual power is computed as 3600/1/interval/pph*1 which in my case (pph=2, interval=10) amounts to 180W. This is the worst case, for even lower power the number of intervals may be any number between 1 and 30.
2. the algorithm averages over 5 samples, however the periods of 5 minutes of virtually zero power are not taken into account.
   This issue does not affect the reported energy consumption which is derived from the counter values.

I think the issue can be resolved by:

1. using the actual time between pulses when computing the actual power (as opposed to the time % 5 minutes);
2. making sure that the samples in the buffer for averaging represent equal time steps (now this is only the case when there is at least 1 pulse per 10 seconds interval).