High frequency mains sensing

In addition to stevenh's answer, I'd like to suggest 2 more alternatives:

1. You can have a separate ADC in each of the floating channels. The ADCs would be linked to the PIC through serial bus (such as SPI). The bus goes to the floating channel through digital optocouplers, which are cheaper than analog optocouplers. Once you have an ADC inside the channel, you can make as many measurements as the ADC has got channels without needing additional optocouplers. This approach was already mentioned in kenny's comment.
2. Each channel could have a dedicated PIC. Each channel with its PIC (and its display, if there are displays) would all float together, and independently of the other channel. You wouldn't need to install a an analog optocoupler for each signal you want to measure. Furthermore, the PICs could talk to each-other through serial link (UART for example). The serial link would go through digital optocouplers.

This is easily done using only 2 bytes of RAM if you implement a 256 times over-sampling filter.

You allocate 2 bytes of RAM to form a 16-bit counter. If you look carefully, you will observe that the upper byte contains the full value of the 16-bit counter / 256.

It is this property that makes this filter so easy.

I'm not at my computer, so I'll try to describe what I mean in pseudo code. Let's call the two bytes that form the counter CntrH & CntrL. As described above, the 8-bit value for the final output is contained in CntrH.

First, subtract the old filter value from the counter.

CntrH : CntrL - CntrH

Now add the 8-bit value from the a/d converter to the 16-bit counter.

A2D + CntrH : CntrL

The filtered output is contained in CntrH.

This is a very slow filter. Therefore, you want to seed the filter upon initialization by taking one a/d sample and loading it into CntrH.

This slow filter means that you want to add new samples fairly quickly. A rough approximation is that your desired filter period takes 256 samples.

In other words, add new samples to the filter at the rate of 10 minutes / 256.

[Edit]

This is extensible by going to a 24-bit accumulator if you want to acquire the samples at a faster rate. This would give you a 2^16 (65536) times over-sampling filter.

Same technique as above but using 3 bytes: CntrH : CntrM : CntrL . As before, the 8-bit filtered output is contained in CntrH.

Given that you want the period to be about 10 minutes (600 seconds), you would accumulate your samples at 600 / (2^16) =~ 9.1ms