I am a beginner at UQ Lab and also new to the forum. I would some advice on developing a surrogate model for quantities derived from a stochastic wind turbine simulator named FAST developed by NREL. I believe there is a module facilitated by UQLab to add the finite element simulator as an external routine and utilize its output for further processing. But I don’t know how and where to start the same. Basically, I have random wind and wave loads that will be described as input to the wind turbine simulator FAST. So, how do I do this for let’s say ‘n’ a number of required runs to achieve an arguably approximate surrogate model for let’s say, blade moments for varying wind speed, turbulence, and wave loads? Any leads shall be helpful. Thanks.
Thanks for your interesting questions. Unfortunately, to my knowledge, there is no literature in the field of uncertainty quantification that deals with this kind of problem. Very high-dimensional, temporally coherent inputs and outputs are a rather rare case for which there is no model in UQLab that inherently supports this. However, we are working on it and it will be part of future releases, but it is unclear when this will be.
Hi there @Subham_Kashyap ,
just to clarify: the issue is not on a programming level, but rather at a much more general theory/research level.
There exist currently no straightforward “black box” surrogate that can take blindly train on a bunch of time series and produce an accurate emulation of the output. This is the research topic of Styfen’s PhD, btw, so he’s definitely an expert on the topic
You can have a look at the topic of physics-informed neural networks, which is probably getting kind of closer to what you are aiming for than anything else I know. But once again, there’s no “one-button” solution available in the literature at the moment.
Well it’s heartening that I have actually encountered a research gap, but what if we process the obtained time series for let’s say 10 minute simulations and further extrapolate the obtained quantiles?
I was hoping to characterise the time series by selecting a quantile level or properties like mean crossing rate for stationary/ non-stationary processes. Also, my ultimate objective remains to map the input to an implicit limit state function to study the design point and further utilise the same for reliability related problems.
Also, thanks for replying and giving me the assurance that atleast I am thinking on the correct lines. Any further discussion shall also help me refine the problem at hand. Eager to hear more.
Hi @Subham_Kashyap,
if what you need is an approximation of 10 minutes statistics on some QoIs as a function of input wind statistics (e.g. maximum load over 10 minutes period for a mean windspeed V and a turbulence level \sigma), you may be willing to have a look to @xujia’s work on stochastic emulators:
As a short follow-up: We recently published a preprint in which we present and demonstrate a method for approximating the dynamic response of complex (e.g., engineering) systems in the time domain. The paper includes two case studies, one of which is the NREL’s 5 MW turbine simulated with OpenFAST, where we predict (among other quantities) the flapwise blade root moment. So it might be of interest to you.
