I have physical model prediction data as well as actual data. From this I can calculate the error of each prediction data point through simple subtraction. I am hoping to train a neural network to be able to assign an error to the input of the physical model.
My current plan is to normalize the error of each data point and assign it as a label to each model input. So the NN would be trained (and validated)on a 1000 data points with the associated error as a label. Once the model is trained I would be able to input one data point and the output of the neural network would be a single class, that is the error. The purpose this would serve would be to tune the physical prediction model.
I have position and velocity data for a satellite. This is going in as an input into an orbit propagation model. The output of this physics model is the position and velocity at some future point in time (2 weeks). This is subtracted from the actual position and velocity at the future point to obtain the error. The error is normalized and assigned to the original input to the orbit propagation model. My goal is to train a model to be able to assign error values (the output) to orbit parameters (the input) and predict what the error would be for any position and velocity vector. Would this kind of architecture work? If so, would you recommend a feedforward or RNN? Thank you.