Regressions
Certainly an interesting discussion.
you are coming at the problem from a certification point of view
Both certification (in that I have been involved in developing AFM performance data in the past) and operations engineering (in that I have spent considerable time in years gone by as an airline operations engineer). As I suggested earlier, there are two distinct areas of interest; takeoff and landing data simulations, where accuracy and precision is necessary for the avoidance of potential legal concerns, and routine performance where the operator is concerned with flight planning and flight following.
multiple different aircraft appendages ..... All of these variants have different performance
As I queried previously, these are significant mods and would, in the usual course of events, be associated with STCs or similar protocols and revised performance data in the AFM/POH supplement(s). Are you suggesting that such did not occur ? Inconceivable within a normally disciplined operation overseen by a competent Regulatory Authority. On that point, of course, you have not identified the State under which the aircraft are registered. The supplements should cover revised takeoff and landing performance but, in all likelihood, may have been a tad relaxed for planning data. there is no usable data in the FOM. suggests the latter to be the case. What is the situation for the takeoff and landing data ?
To do this
That's all fine and, philosophically, no different to what legions of airline operations engineers do for their daily crust.
Also for survey
Likewise. It is acknowledged that lighties often have a few gaps in the area of flight planning data.
I don't see why there should be a "risk"
The risk with which I was always concerned related to the degree to which the model might be unpredictable in areas of the envelope, the probability increasing with order. This is particularly the case with multivariate analyses where the subsequent interpolation requirements require significant development testing to ensure that the thing is well behaved. Very early on, I came to the decision that a safer strategy was to stick to lower orders and, where necessary, just segment the data sets to achieve whatever level of precision and accuracy I was after. From similar reasoning, I abandoned multivariate analyses and used low order interpolation routines which I knew weren't going to bite me on the tail in operational use.
as long as you do not attempt to extrapolate outside of your data set.
Forcing the use of low orders, associated with limited manual extrapolation for development, provided the facility to extrapolate in operations to determine non-critical limitations which, on occasion, were useful for assessment of operational penalties for degraded systems operation.
... I use fourth order ... I am using fifth order for some of the other parts of the model
I'd avoid that like the plague but let's just agree to disagree on preferred strategies ?
I am not using my calculations for certifications
I see no fundamental difference - the potential problem relates to model discipline in use. The higher the order the higher the testing load.
As long as you stay within your data set, higher orders greatly increase accuracy.
So long as the data set is significantly large - I just don't see any real advantage in reducing the number of data sets to a minimum when it is so easy to set up additional lower order models to avoid the problem altogether.
near the stall would be really inaccurate if I did not use high orders.
By segmenting the datasets the problem doesn't arise