My preferred method (when using a DSLR) for capturing flats was to only pay attention to the appropriate channel(s) - Red for Ha, Blue and Green for OIII. So, when capturing flats for Ha, I only pay attention to the exposure for my red channel. So, in your case, I'd pay attention the the 20,000 ADU peak for the red and ignore the 100 - 200 blips for each of the green and blue channels.
When you calibrate your data, your calibration software SHOULD be calibrating each light pixel based upon it's corresponding pixel in your calibration frames. All of these files (lights and calibration files) are still monochrome at this point. As such, if I'm using an Ha filter, then a red pixel in my light will be calibrated with the corresponding red pixel in my flat (and dark, bias, etc.). Given this, your calibrated light should be correct. The only difference from a broadband image is that the data (for an Ha image, as an example) in your blue and green channels is just useless noise.
The next step is the difference between calibrating broadband versus narrowband. I do NOT debayer my calibrated lights. This is because I don't want any debayering algorithm to try to estimate the values of each of my color pixels (as is normally done with debayering). What I want is the useful data from the narrowband data. So, I'll extract the red only pixels and end up with an image that has 1/2 the dimensions of my original light. From that point, I normalize, align, and stack these 1/2 dimension images.
For Oiii, I extract the green and blue pixels and process the same as red pixels for Ha. However, in this case, my green data has twice the SNR as the blue pixels. So, when I have each of my stacked images for blue and green, I'll then copy the green channel and the stack the 2 greens plus blue using an average for these images. This maintains the correct SNR ratios between the green and blue pixels.
Perhaps more complicated, but the results should be accurate. Hope this helps.