I'm going to highlight updates to the original description in italic and add developments to the bottom. 

It's currently "working" in so far as it returns numbers:      https://thingspeak.com/channels/1098352     (the data feed is erratic at the moment as I'm working on it from time to time).

It survives overnight from the fridge and freezer current and even continues to transmit. The capacitor must have low enough value to give measurable voltage changes at low power use, while large enough to maintain operation through a power cut. 

However the numbers returned are not accurate. I need to refine the coefficients in my formula that calculates the power drain. I have removed the CT input and manually charged the capacitor then used its voltage to make reasonable estimates of the power drain and now intend to capture real data that will let me solve more accurately for the coefficients. 

I was surprised to find my 8Mhz Pro Mini operates below 1.8v and the cheap 433Mhz RF transmits at below 2v, going through several walls. However I aim to operate it at above 2v to leave a margin for surviving a power cut.  I use the RCSwitch library and tried each protocol to find the ones with lowest energy use.  A check digit is necessary as signals are occasionally corrupted. I used avrdude to change the brown out detection fuse. I intended to reduce the frequency to 1MHz as well but that hasn't been necessary and is harder than changing the BoD.

An ESP8266 is currently receiving the 433RF and updating the online thingspeak log.

The original aim of the project was to reveal if any appliance was left on at the end of the day and this has been achieved. It can't detect an 11 w low energy bulb but a cooker hotplate or anything significant now shows up. Trying to expand the function to give a reasonable overall power use reading and is too tempting to ignore. As with commercial clip on power monitors there's inaccuracy due to the supply voltage varying slightly from 230v - the power depends on the current and voltage.  

Calibration Stage 1

There is current drain from:

a) being awake 

b) deep sleep

c) transmitting

d) start and end voltage

e) I don't think I can ignore capacitor leakage so a futher test is required. Initial measurements suggest as high as 15uA which surprises me. 

Counting the time or number for each of these for a period with known energy consumption (from the meter reading) will allow solving for the unknown current drain from each. There are 4 unknowns so in theory 4 sets of data will suffice, but in practice periods will often be similar to each other and I need very different sampling periods. I'm coding it to report a - d every hour and I will read the meter hourly.

The start and end voltage will have little effect as the small difference is averaged over the hour.

However it's not quite that simple. The current drain will also depend on the voltage. For example transmit current may vary with the square of the voltage. However the code attempts to keep the voltage  in a narrow range so the variation due to voltage may be small. If the accuracy of this stage is insufficient.....

Stage 2

To measure the variation with voltage the CT will be removed and the capacitor charged manually to 4v (the transmitter seems unaffected by brief bursts above its 3.6v spec max. The reported data can then be used to fit the variation with V. In this case the changes in V will be significant since the end V will be at the limit of transmit operation. 

Stage 3

If stage 2 does not finalise the parameters, a period of controlled power use will be monitored. All appliances will be disconnected and then a known power drain (such as a 1 kw electric bar) turned on.

This will be repeated with 2 kw then 3 kw.

The meter shows divisions of 0.02 kw so it won't be necessary to run the test for as long as an hour. 

=====================================================================================

Have some data now and calculated  the coefficients values ( for current drain during each operation).

It's not so hard fitting the coefficients because some operations (like sampling the voltage) take very little current so that varying the transmit and awake values are the only way to fit close to observation. 

Different periods are mostly giving similar results but the occasional outlier which may be due to fault reading of the meter or else a fault in my method. So more logging is needed. 

Also wondering if I can correlate the harvested current with the mains current. I have discovered that CT probaby has 2000 turns, so mains current C should yield C/2000 in the monitor. I previously tried measuring the yield using an ammeter but not sure how to convert the AC current into DC coming out my rectification/voltage double. I presume it's halved and then minor losses in the two capacitors and Schotty diodes. I was also unsure what happens if the voltage from the CT is lower than that in the supercapacitor but now suspect a CT has to unload it's energy so the voltage increases until it can discharge it. Latest data supports a reduction of 1/4000 but early days yet. 

Changes needed.

Transmission of data. I keep the packet small (24 bit number) which has to include a check digit, a sequence number, voltage, and power estimate. I have a max of 162 for the value of power. Since I don't need the same resolution at high power as low power I'm going to transmit as the log or the square root of the power. 

Voltage limitation.

I now realise from the figures that if many appliances were on at once the Pro Mini could not keep the capacitor voltage below the Zener diode point and so the readings would be incorrect. I will need to code for dummy transmissions to use up power in the event the voltage gets high, though this should be unusual and brief. (Thingspeak only allows update at most every 15 seconds so no point transmitting valid readings more frequently.)

Problems

1) The voltage on the capacitor as measured by the Pro Mini jumps about a little, but a little is enough at very low power use, to produce readings such as 50W, 100W, 50W, 0W. These are caused by a variation in the least significant digit of voltage. There may be slight ripple on the capacitor voltage but enough to change the voltage just enough to change the last digit at the exact moment I sample the voltage.

I will have to think about averaging, either over several intervals (currently 20s) or else sample more frequently and average. Sampling is cheap but it has to be awake to do so which uses up precious charge and this problem arises at very low power use when conserving charge is important. Still it's reporting more than once/minute through the night so it's feasible.

2) The power calculation is producing exaggerated results. I am beginning to think the current yield from the CT does vary according to the voltage on the capacitor. I can fix this with a simple scaling adjustment once I estimate what it is.

Improving the logging is taking time. Bugs creep in. Seems I’m good at the big picture but not so with the tiny details which have to be spot on for the code to work as I want. Mind you I have to get the code talking between 2 x ESP8266 and the Pro Mini. I can’t bench test as it has to run off the supercapacitor for meaningful results, so RF logging is necessary. Then I have to read the meter, run it for a while and read the meter again. 

==============================================================================

Latest problem: to understand the yield from CT, and model the circuit current draw, I need to report the number of transmits, read voltage,  time asleep, awake, and the cap voltage, but there isn’t always the energy in the capacitor for all the transmitting and time awake to separate the transmissions for reliable receipt. Furthermore the pro mini then has additional function distorting the results. Looks like I will have to connect a powered esp8266 to do the transmitting and let the pro mini just do as it was meant to, albeit with a little serial communication. I need to finish off another project so suspending this project for a few weeks. Later thought - can't interface the esp8266 at 3.3v to Pro Mini at ~2.1v. Another plan required: Powered ESP8266 on its own using GPIOs and resistors/diodes to control the capacitor voltage.

==================================================================================

ESP8266 code debugged and power bank discarded in favour of a mains power supply - took a while to work out how to route it safely to the external electricity cupboard.

Now obtaining useful looking data. However I realise I now need to learn about "multiple regression" in order to find coefficients: I have 29 voltage levels I monitor in 0.1v steps from 1.8v to 4.7v in order to find the yield from the CT at each level. I know for each level the voltage and the resistance, hence the current. There's a relationship between the current I measure at each level and the total power use as measured by the meter. So I have an equation with 29 unknowns which I need to solve. Hopefully I will discern a pattern as they will all be related in some way.

I can see that my earlier belief that my RF transmissions use very few milliamps (2-5 at 2v) appears correct as the yield during the night is very low yet I was getting regular transmissions from my Pro Mini throughout when I was testing it.