In any implementation of an ADC, there's going to be some error to the measurement. In many cases, the error is primarily due to the resolution of the ADC so it's safe to assume that the error in the measurement is plus or minus half the ADC resolution. However, in this project, I'm hoping to get the highest possible resolution from the system, so a more careful analysis of sources of error is appropriate.

Here's the specifications of my large 200kg load cells from aliexpress. 

This suggests that I'll get about 6mV of signal with an excitation voltage of 3.0V. The HX711 spec sheet doesn't mention a noise floor, but assuming there's no noise on the board, the 6mV signal at maximum load corresponds to a useful 15 bits of resolution. This suggests a resolution of roughly 0.01 kg! However, in reality, there IS noise on the circuit, and there are other, much more significant sources of error! The HX711 ADC chip does include 128x amplification of the signal, but as this will amplify noise as well, and I know most 22-24 bit ADCs are noise limited around 20-21 bits, I doubt the amplification is helping with resolution (although it may be helping to keep chip cost low).

I don't have a good specification for noise on this circuit, but a quick test suggested that the repeatability of the system is only around +/- 0.5 kg. I'm hoping this will improve when I have all the connections properly soldered instead of connected with probes and alligator clips, but I'm also exciting the load cells with 5V right now, so the decrease to 3V will make any electrical noise more significant. +/- 0.5 kg is only about 8.5 bits of useful resolution, and I'm honestly disappointed with this result (my stretch goal is 0.1 kg resolution) but between the ADC board sourced on Ebay and the low-cost postal scale load cells, I won't be surprised if I can't improve much on this. While it's disappointing, 0.5 kg resolution is plenty for the primary goal of measuring nectar flows even if it doesn't reliably show more subtle effects like foraging bees leaving in the morning and returning at night!

There are two other sources of error that I'm concerned about and won't be able to measure until the system is completed. First, the above specification claims a 0.025% of full scale change in both zero and output per 10 degrees Celsius. That's actually really significant, as I expect to see roughly 100 degrees of temperature change over a year! Now a typical summer day day will see more like 20 degrees of change, but temperature could potentiall shift the reading by 0.1 kg per 10 degrees. Further, I know amplification circuits can be quite temperature sensitive, so I expect a similar, if not greater error introduced by the amplification on the HX711 chip.

I should be able to remove some of the temperature error by calibrating the scale (assuming the temperature sensitivity does not drift) but an even worse source of error may be in long-term drift of the load cells. The specification above claims around 0.05 kg of drift over 30 minutes, but the error will continue to accumulate after that 30 minutes. I'm honestly not sure to expect, although I am fairly confident that the error will stabilize after a few days and won't significantly impact measurement of nectar collection that could be a kilogram per day or more.

There's a lot of unknowns, but my first rough test of repeatability shows that the system will work for my main goal of tracking nectar flows (even if it doesn't have the higher resolution I'd hoped for) and I'll be able to work on future designs to try to reduce the error further.