polar and ivr wls on case3_unbalance "locally_infeasible" but actually not

[MartaVanin]

Look into ipopt settings: ipopt returns a "locally_infeasible" exit status but also a correct solution.

[timmyfaraday]

Is this always the case? I only got such behavior when really pushing down the variance of the uncertainty. The documentation suggests looking into a different starting point. The exploitation of the starting point given by a power flow is also something we need to look into, in my opinion.

[MartaVanin]

it does not happen with higher variances, but at that point the errors I generate on the fake measurement are so high that the estimation error is not negligible anymore. One thing we could do is replace the weight in the residual definition, using instead of the (square of the) variance ,100 times its amount. I think it is acceptable to do so and it won't change the result, as the ratios between the weights of more/less accurate measurements won't change, but I am not sure the literature supports this approach, that needs to be double checked. The starting point is absolutely something to discuss, we should choose an option between:

• flat start. Very neutral, but computationally not the best.
• using the result of the powerflow. Good starting point computationally, but we'd need to justify it: how would you have that in real-life?
• using the measurements as they are (with their errors). This might make your life less fun if you have bad measurements, but I think it is acceptable for our work right now.
• other suggestions..

[timmyfaraday]

I'm not sure changing the expression of the residual is such a good idea. Additionally, a possible other startpoint could be the result of the preceeding state estimation given a number of them are performed in sequence.

[MartaVanin]

I added a rescaler possibility to the weights. If not explicitly declared otherwise, it defaults to 1, so the expression of the residuals remains unchanged. We can discuss this on Wednesday. Indeed, if you do SE for two succeeding time steps I think what you suggested is possibly the best approach. I reckon a better starting point than that could only (maybe) be achieved via statistical/machine learning techniques.

[MartaVanin]

old issue, these problems seem to be gone with the enwl dataset and the recent changes. I'll just close it