@bubuche There is a grade lag at the start of a descent due to smoothing. So the calculation uses the previous flat section (0% grade) until the smoothing catches up. And even then you might be at -8% but the calculation uses for ex. -2%. At this speed, you are going above the assumed terminal velocity (which is too low), and get high power spikes,
Without a cadence sensor, I could zero the power when going downhill and avoid these short spikes.
So replace a false positive (what you see) with a false negative.
I guess the false positive is the worse issue since it impacts NP, averages, etc.
So this needs fixing, but without breaking other cases. I’ll have a think.